rag-architect — Skillopedia

RAG Architect - POWERFUL Overview The RAG (Retrieval-Augmented Generation) Architect skill provides comprehensive tools and knowledge for designing, implementing, and optimizing production-grade RAG pipelines. This skill covers the entire RAG ecosystem from document chunking strategies to evaluation frameworks, enabling you to build scalable, efficient, and accurate retrieval systems. Core Competencies 1. Document Processing & Chunking Strategies Fixed-Size Chunking - Character-based chunking : Simple splitting by character count (e.g., 512, 1024, 2048 chars) - Token-based chunking : Splittin…

, re.MULTILINE), # Markdown headers\n re.compile(r'^(.+)\\n[=-]+\\s*

Important: agents should read /llm.txt, /llms.txt, or /.well-known/skills.json to discover the public Skillopedia API.

, re.MULTILINE), # Underlined headers\n re.compile(r'^\\d+\\.\\s*(.+)

Important: agents should read /llm.txt, /llms.txt, or /.well-known/skills.json to discover the public Skillopedia API.

, re.MULTILINE), # Numbered sections\n ]\n \n def chunk(self, text: str) -> List[Dict[str, Any]]:\n sections = self._identify_sections(text)\n chunks = []\n chunk_id = 0\n \n for section in sections:\n section_chunks = self._chunk_section(section, chunk_id)\n chunks.extend(section_chunks)\n chunk_id += len(section_chunks)\n \n return chunks\n \n def _identify_sections(self, text: str) -> List[Dict[str, Any]]:\n \"\"\"Identify sections based on headings.\"\"\"\n sections = []\n lines = text.split('\\n')\n current_section = {'heading': 'Introduction', 'content': '', 'level': 0}\n \n for line in lines:\n is_heading = False\n heading_level = 0\n heading_text = line.strip()\n \n # Check for markdown headers\n if line.strip().startswith('#'):\n level = len(line) - len(line.lstrip('#'))\n if level \u003c= 6:\n heading_text = line.strip('#').strip()\n heading_level = level\n is_heading = True\n \n # Check for underlined headers \n elif len(sections) > 0 and line.strip() and all(c in '=-' for c in line.strip()):\n # Previous line might be heading\n if current_section['content']:\n content_lines = current_section['content'].strip().split('\\n')\n if content_lines:\n potential_heading = content_lines[-1].strip()\n if len(potential_heading) > 0 and len(potential_heading) \u003c 100:\n # Treat as heading\n current_section['content'] = '\\n'.join(content_lines[:-1])\n sections.append(current_section)\n current_section = {\n 'heading': potential_heading,\n 'content': '',\n 'level': 1 if '=' in line else 2\n }\n continue\n \n if is_heading:\n if current_section['content'].strip():\n sections.append(current_section)\n current_section = {\n 'heading': heading_text,\n 'content': '',\n 'level': heading_level\n }\n else:\n current_section['content'] += line + '\\n'\n \n # Add final section\n if current_section['content'].strip():\n sections.append(current_section)\n \n return sections\n \n def _chunk_section(self, section: Dict[str, Any], start_id: int) -> List[Dict[str, Any]]:\n \"\"\"Chunk a single section.\"\"\"\n content = section['content'].strip()\n if not content:\n return []\n \n heading = section['heading']\n chunks = []\n \n # If section is small enough, return as single chunk\n if len(content) \u003c= self.max_size:\n chunks.append({\n 'id': start_id,\n 'text': f\"{heading}\\n\\n{content}\" if heading else content,\n 'start': 0,\n 'end': len(content),\n 'size': len(content),\n 'heading': heading,\n 'level': section['level']\n })\n return chunks\n \n # Split large sections by paragraphs\n paragraphs = [p.strip() for p in content.split('\\n\\n') if p.strip()]\n current_chunk = []\n current_size = len(heading) + 2 if heading else 0 # Account for heading\n chunk_id = start_id\n \n for paragraph in paragraphs:\n paragraph_size = len(paragraph)\n \n if current_size + paragraph_size > self.max_size and current_chunk:\n # Save current chunk\n chunk_text = '\\n\\n'.join(current_chunk)\n if heading and chunk_id == start_id:\n chunk_text = f\"{heading}\\n\\n{chunk_text}\"\n \n chunks.append({\n 'id': chunk_id,\n 'text': chunk_text,\n 'start': 0,\n 'end': len(chunk_text),\n 'size': len(chunk_text),\n 'heading': heading if chunk_id == start_id else f\"{heading} (continued)\",\n 'level': section['level']\n })\n chunk_id += 1\n current_chunk = [paragraph]\n current_size = paragraph_size\n else:\n current_chunk.append(paragraph)\n current_size += paragraph_size + 2 # Account for newlines\n \n # Add final chunk\n if current_chunk:\n chunk_text = '\\n\\n'.join(current_chunk)\n if heading and chunk_id == start_id:\n chunk_text = f\"{heading}\\n\\n{chunk_text}\"\n elif heading:\n chunk_text = f\"{heading} (continued)\\n\\n{chunk_text}\"\n \n chunks.append({\n 'id': chunk_id,\n 'text': chunk_text,\n 'start': 0,\n 'end': len(chunk_text),\n 'size': len(chunk_text),\n 'heading': heading if chunk_id == start_id else f\"{heading} (continued)\",\n 'level': section['level']\n })\n \n return chunks\n\n\nclass ChunkAnalyzer:\n \"\"\"Analyzes chunks and provides quality metrics.\"\"\"\n \n def __init__(self):\n self.vocabulary = set()\n self.word_freq = Counter()\n \n def analyze_chunks(self, chunks: List[Dict[str, Any]]) -> Dict[str, Any]:\n \"\"\"Comprehensive chunk analysis.\"\"\"\n if not chunks:\n return {'error': 'No chunks to analyze'}\n \n sizes = [chunk['size'] for chunk in chunks]\n \n # Basic size statistics\n size_stats = {\n 'count': len(chunks),\n 'mean': statistics.mean(sizes),\n 'median': statistics.median(sizes),\n 'std': statistics.stdev(sizes) if len(sizes) > 1 else 0,\n 'min': min(sizes),\n 'max': max(sizes),\n 'total': sum(sizes)\n }\n \n # Boundary quality analysis\n boundary_quality = self._analyze_boundary_quality(chunks)\n \n # Semantic coherence (simple heuristic)\n coherence_score = self._calculate_semantic_coherence(chunks)\n \n # Vocabulary distribution\n vocab_stats = self._analyze_vocabulary(chunks)\n \n return {\n 'size_statistics': size_stats,\n 'boundary_quality': boundary_quality,\n 'semantic_coherence': coherence_score,\n 'vocabulary_statistics': vocab_stats\n }\n \n def _analyze_boundary_quality(self, chunks: List[Dict[str, Any]]) -> Dict[str, Any]:\n \"\"\"Analyze how well chunks respect natural boundaries.\"\"\"\n sentence_breaks = 0\n word_breaks = 0\n total_chunks = len(chunks)\n \n sentence_endings = re.compile(r'[.!?]\\s*

Important: agents should read /llm.txt, /llms.txt, or /.well-known/skills.json to discover the public Skillopedia API.

)\n \n for chunk in chunks:\n text = chunk['text'].strip()\n if not text:\n continue\n \n # Check if chunk ends with sentence boundary\n if sentence_endings.search(text):\n sentence_breaks += 1\n \n # Check if chunk ends with word boundary\n if text[-1].isalnum() or text[-1] in '.!?':\n word_breaks += 1\n \n return {\n 'sentence_boundary_ratio': sentence_breaks / total_chunks if total_chunks > 0 else 0,\n 'word_boundary_ratio': word_breaks / total_chunks if total_chunks > 0 else 0,\n 'clean_breaks': sentence_breaks,\n 'total_chunks': total_chunks\n }\n \n def _calculate_semantic_coherence(self, chunks: List[Dict[str, Any]]) -> float:\n \"\"\"Simple semantic coherence heuristic based on vocabulary overlap.\"\"\"\n if len(chunks) \u003c 2:\n return 1.0\n \n coherence_scores = []\n \n for i in range(len(chunks) - 1):\n chunk1_words = set(re.findall(r'\\b\\w+\\b', chunks[i]['text'].lower()))\n chunk2_words = set(re.findall(r'\\b\\w+\\b', chunks[i+1]['text'].lower()))\n \n if not chunk1_words or not chunk2_words:\n continue\n \n # Jaccard similarity as coherence measure\n intersection = len(chunk1_words & chunk2_words)\n union = len(chunk1_words | chunk2_words)\n \n if union > 0:\n coherence_scores.append(intersection / union)\n \n return statistics.mean(coherence_scores) if coherence_scores else 0.0\n \n def _analyze_vocabulary(self, chunks: List[Dict[str, Any]]) -> Dict[str, Any]:\n \"\"\"Analyze vocabulary distribution across chunks.\"\"\"\n all_words = []\n chunk_vocab_sizes = []\n \n for chunk in chunks:\n words = re.findall(r'\\b\\w+\\b', chunk['text'].lower())\n all_words.extend(words)\n chunk_vocab_sizes.append(len(set(words)))\n \n total_vocab = len(set(all_words))\n word_freq = Counter(all_words)\n \n return {\n 'total_vocabulary': total_vocab,\n 'avg_chunk_vocabulary': statistics.mean(chunk_vocab_sizes) if chunk_vocab_sizes else 0,\n 'vocabulary_diversity': total_vocab / len(all_words) if all_words else 0,\n 'most_common_words': word_freq.most_common(10)\n }\n\n\nclass ChunkingOptimizer:\n \"\"\"Main optimizer that tests different chunking strategies.\"\"\"\n \n def __init__(self):\n self.analyzer = ChunkAnalyzer()\n \n def optimize(self, corpus: DocumentCorpus, config: Dict[str, Any] = None) -> Dict[str, Any]:\n \"\"\"Test all chunking strategies and recommend the best one.\"\"\"\n config = config or {}\n \n strategies = self._create_strategies(config)\n results = {}\n \n print(f\"Testing {len(strategies)} chunking strategies...\")\n \n for strategy in strategies:\n print(f\" Testing {strategy.name}...\")\n strategy_results = self._test_strategy(corpus, strategy)\n results[strategy.name] = strategy_results\n \n # Recommend best strategy\n recommendation = self._recommend_strategy(results)\n \n return {\n 'corpus_info': {\n 'document_count': len(corpus.documents),\n 'total_size': sum(d['size'] for d in corpus.documents),\n 'avg_document_size': statistics.mean([d['size'] for d in corpus.documents])\n },\n 'strategy_results': results,\n 'recommendation': recommendation,\n 'sample_chunks': self._generate_sample_chunks(corpus, recommendation['best_strategy'])\n }\n \n def _create_strategies(self, config: Dict[str, Any]) -> List[ChunkingStrategy]:\n \"\"\"Create all chunking strategies to test.\"\"\"\n strategies = []\n \n # Fixed-size strategies\n for size in config.get('fixed_sizes', [512, 1000, 1500]):\n for overlap in config.get('overlaps', [50, 100]):\n strategies.append(FixedSizeChunker(size, overlap, 'char'))\n \n # Sentence-based strategies\n for max_size in config.get('sentence_max_sizes', [800, 1200]):\n strategies.append(SentenceChunker(max_size))\n \n # Paragraph-based strategies\n for max_size in config.get('paragraph_max_sizes', [1500, 2000]):\n strategies.append(ParagraphChunker(max_size))\n \n # Semantic strategies\n for max_size in config.get('semantic_max_sizes', [1200, 1800]):\n strategies.append(SemanticChunker(max_size))\n \n return strategies\n \n def _test_strategy(self, corpus: DocumentCorpus, strategy: ChunkingStrategy) -> Dict[str, Any]:\n \"\"\"Test a single chunking strategy.\"\"\"\n all_chunks = []\n document_results = []\n \n for doc in corpus.documents:\n try:\n chunks = strategy.chunk(doc['content'])\n all_chunks.extend(chunks)\n \n doc_analysis = self.analyzer.analyze_chunks(chunks)\n document_results.append({\n 'path': doc['path'],\n 'chunk_count': len(chunks),\n 'analysis': doc_analysis\n })\n except Exception as e:\n print(f\" Error processing {doc['path']}: {e}\")\n continue\n \n # Overall analysis\n overall_analysis = self.analyzer.analyze_chunks(all_chunks)\n \n return {\n 'strategy_config': strategy.config,\n 'total_chunks': len(all_chunks),\n 'overall_analysis': overall_analysis,\n 'document_results': document_results,\n 'performance_score': self._calculate_performance_score(overall_analysis)\n }\n \n def _calculate_performance_score(self, analysis: Dict[str, Any]) -> float:\n \"\"\"Calculate overall performance score for a strategy.\"\"\"\n if 'error' in analysis:\n return 0.0\n \n size_stats = analysis['size_statistics']\n boundary_quality = analysis['boundary_quality']\n coherence = analysis['semantic_coherence']\n \n # Normalize metrics to 0-1 range and combine\n size_consistency = 1.0 - min(size_stats['std'] / size_stats['mean'], 1.0) if size_stats['mean'] > 0 else 0\n boundary_score = (boundary_quality['sentence_boundary_ratio'] + boundary_quality['word_boundary_ratio']) / 2\n coherence_score = coherence\n \n # Weighted combination\n return (size_consistency * 0.3 + boundary_score * 0.4 + coherence_score * 0.3)\n \n def _recommend_strategy(self, results: Dict[str, Any]) -> Dict[str, Any]:\n \"\"\"Recommend the best chunking strategy based on analysis.\"\"\"\n best_strategy = None\n best_score = 0\n \n strategy_scores = {}\n \n for strategy_name, result in results.items():\n score = result['performance_score']\n strategy_scores[strategy_name] = score\n \n if score > best_score:\n best_score = score\n best_strategy = strategy_name\n \n return {\n 'best_strategy': best_strategy,\n 'best_score': best_score,\n 'all_scores': strategy_scores,\n 'reasoning': self._generate_reasoning(best_strategy, results[best_strategy] if best_strategy else None)\n }\n \n def _generate_reasoning(self, strategy_name: str, result: Dict[str, Any]) -> str:\n \"\"\"Generate human-readable reasoning for the recommendation.\"\"\"\n if not result:\n return \"No valid strategy found.\"\n \n analysis = result['overall_analysis']\n size_stats = analysis['size_statistics']\n boundary = analysis['boundary_quality']\n \n reasoning = f\"Recommended '{strategy_name}' because:\\n\"\n reasoning += f\"- Average chunk size: {size_stats['mean']:.0f} characters\\n\"\n reasoning += f\"- Size consistency: {size_stats['std']:.0f} std deviation\\n\"\n reasoning += f\"- Boundary quality: {boundary['sentence_boundary_ratio']:.2%} clean sentence breaks\\n\"\n reasoning += f\"- Semantic coherence: {analysis['semantic_coherence']:.3f}\\n\"\n \n return reasoning\n \n def _generate_sample_chunks(self, corpus: DocumentCorpus, strategy_name: str) -> List[Dict[str, Any]]:\n \"\"\"Generate sample chunks using the recommended strategy.\"\"\"\n if not strategy_name or not corpus.documents:\n return []\n \n # Create strategy instance\n strategy = None\n if 'fixed_size' in strategy_name:\n strategy = FixedSizeChunker()\n elif 'sentence' in strategy_name:\n strategy = SentenceChunker()\n elif 'paragraph' in strategy_name:\n strategy = ParagraphChunker()\n elif 'semantic' in strategy_name:\n strategy = SemanticChunker()\n \n if not strategy:\n return []\n \n # Get chunks from first document\n sample_doc = corpus.documents[0]\n chunks = strategy.chunk(sample_doc['content'])\n \n # Return first 3 chunks as samples\n return chunks[:3]\n\n\ndef main():\n \"\"\"Main function with command-line interface.\"\"\"\n parser = argparse.ArgumentParser(description='Analyze documents and recommend optimal chunking strategy')\n parser.add_argument('directory', help='Directory containing text/markdown documents')\n parser.add_argument('--output', '-o', help='Output file for results (JSON format)')\n parser.add_argument('--config', '-c', help='Configuration file (JSON format)')\n parser.add_argument('--extensions', nargs='+', default=['.txt', '.md', '.markdown'], \n help='File extensions to process')\n parser.add_argument('--verbose', '-v', action='store_true', help='Verbose output')\n \n args = parser.parse_args()\n \n # Load configuration\n config = {}\n if args.config and os.path.exists(args.config):\n with open(args.config, 'r') as f:\n config = json.load(f)\n \n try:\n # Load corpus\n print(f\"Loading documents from {args.directory}...\")\n corpus = DocumentCorpus(args.directory, args.extensions)\n \n # Run optimization\n optimizer = ChunkingOptimizer()\n results = optimizer.optimize(corpus, config)\n \n # Save results\n if args.output:\n with open(args.output, 'w') as f:\n json.dump(results, f, indent=2)\n print(f\"Results saved to {args.output}\")\n \n # Print summary\n print(\"\\n\" + \"=\"*60)\n print(\"CHUNKING OPTIMIZATION RESULTS\")\n print(\"=\"*60)\n \n corpus_info = results['corpus_info']\n print(f\"Corpus: {corpus_info['document_count']} documents, {corpus_info['total_size']:,} characters\")\n \n recommendation = results['recommendation']\n print(f\"\\nRecommended Strategy: {recommendation['best_strategy']}\")\n print(f\"Performance Score: {recommendation['best_score']:.3f}\")\n print(f\"\\nReasoning:\\n{recommendation['reasoning']}\")\n \n if args.verbose:\n print(\"\\nAll Strategy Scores:\")\n for strategy, score in recommendation['all_scores'].items():\n print(f\" {strategy}: {score:.3f}\")\n \n print(\"\\nSample Chunks:\")\n for i, chunk in enumerate(results['sample_chunks'][:2]):\n print(f\"\\nChunk {i+1} ({chunk['size']} chars):\")\n print(\"-\" * 40)\n print(chunk['text'][:200] + \"...\" if len(chunk['text']) > 200 else chunk['text'])\n \n except Exception as e:\n print(f\"Error: {e}\")\n return 1\n \n return 0\n\n\nif __name__ == '__main__':\n exit(main())","content_type":"text/x-python; charset=utf-8","language":"python","size":29512,"content_sha256":"390cf0b370e6cabbcdc367f809e455c36da4651d7fa2808938b2a958e416d0cc"},{"filename":"rag_pipeline_designer.py","content":"#!/usr/bin/env python3\n\"\"\"\nRAG Pipeline Designer - Designs complete RAG pipelines based on requirements.\n\nThis script analyzes requirements and generates a comprehensive RAG pipeline design\nincluding architecture diagrams, component recommendations, configuration templates,\nand cost projections.\n\nComponents designed:\n- Chunking strategy recommendation\n- Embedding model selection\n- Vector database recommendation \n- Retrieval approach (dense/sparse/hybrid)\n- Reranking configuration\n- Evaluation framework setup\n- Production deployment patterns\n\nNo external dependencies - uses only Python standard library.\n\"\"\"\n\nimport argparse\nimport json\nimport math\nimport os\nfrom typing import Dict, List, Tuple, Any, Optional\nfrom dataclasses import dataclass, asdict\nfrom enum import Enum\n\n\nclass Scale(Enum):\n \"\"\"System scale categories.\"\"\"\n SMALL = \"small\" # \u003c 1M documents, \u003c 1K queries/day\n MEDIUM = \"medium\" # 1M-100M documents, 1K-100K queries/day \n LARGE = \"large\" # 100M+ documents, 100K+ queries/day\n\n\nclass DocumentType(Enum):\n \"\"\"Document type categories.\"\"\"\n TEXT = \"text\" # Plain text, articles\n TECHNICAL = \"technical\" # Documentation, manuals\n CODE = \"code\" # Source code files\n SCIENTIFIC = \"scientific\" # Research papers, journals\n LEGAL = \"legal\" # Legal documents, contracts\n MIXED = \"mixed\" # Multiple document types\n\n\nclass Latency(Enum):\n \"\"\"Latency requirements.\"\"\"\n REAL_TIME = \"real_time\" # \u003c 100ms\n INTERACTIVE = \"interactive\" # \u003c 500ms \n BATCH = \"batch\" # > 1s acceptable\n\n\n@dataclass\nclass Requirements:\n \"\"\"RAG system requirements.\"\"\"\n document_types: List[str]\n document_count: int\n avg_document_size: int # characters\n queries_per_day: int\n query_patterns: List[str] # e.g., [\"factual\", \"conversational\", \"analytical\"]\n latency_requirement: str\n budget_monthly: float # USD\n accuracy_priority: float # 0-1 scale\n cost_priority: float # 0-1 scale\n maintenance_complexity: str # \"low\", \"medium\", \"high\"\n\n\n@dataclass\nclass ComponentRecommendation:\n \"\"\"Recommendation for a pipeline component.\"\"\"\n name: str\n type: str\n config: Dict[str, Any]\n rationale: str\n pros: List[str]\n cons: List[str]\n cost_monthly: float\n\n\n@dataclass\nclass PipelineDesign:\n \"\"\"Complete RAG pipeline design.\"\"\"\n chunking: ComponentRecommendation\n embedding: ComponentRecommendation \n vector_db: ComponentRecommendation\n retrieval: ComponentRecommendation\n reranking: Optional[ComponentRecommendation]\n evaluation: ComponentRecommendation\n total_cost: float\n architecture_diagram: str\n config_templates: Dict[str, Any]\n\n\nclass RAGPipelineDesigner:\n \"\"\"Main pipeline designer class.\"\"\"\n \n def __init__(self):\n self.embedding_models = self._load_embedding_models()\n self.vector_databases = self._load_vector_databases()\n self.chunking_strategies = self._load_chunking_strategies()\n \n def design_pipeline(self, requirements: Requirements) -> PipelineDesign:\n \"\"\"Design complete RAG pipeline based on requirements.\"\"\"\n print(f\"Designing RAG pipeline for {requirements.document_count:,} documents...\")\n \n # Determine system scale\n scale = self._determine_scale(requirements)\n print(f\"System scale: {scale.value}\")\n \n # Design each component\n chunking = self._recommend_chunking(requirements, scale)\n embedding = self._recommend_embedding(requirements, scale) \n vector_db = self._recommend_vector_db(requirements, scale)\n retrieval = self._recommend_retrieval(requirements, scale)\n reranking = self._recommend_reranking(requirements, scale)\n evaluation = self._recommend_evaluation(requirements, scale)\n \n # Calculate total cost\n total_cost = (chunking.cost_monthly + embedding.cost_monthly + \n vector_db.cost_monthly + retrieval.cost_monthly + \n evaluation.cost_monthly)\n if reranking:\n total_cost += reranking.cost_monthly\n \n # Generate architecture diagram\n architecture = self._generate_architecture_diagram(\n chunking, embedding, vector_db, retrieval, reranking, evaluation\n )\n \n # Generate configuration templates\n configs = self._generate_config_templates(\n chunking, embedding, vector_db, retrieval, reranking, evaluation\n )\n \n return PipelineDesign(\n chunking=chunking,\n embedding=embedding,\n vector_db=vector_db,\n retrieval=retrieval,\n reranking=reranking,\n evaluation=evaluation,\n total_cost=total_cost,\n architecture_diagram=architecture,\n config_templates=configs\n )\n \n def _determine_scale(self, req: Requirements) -> Scale:\n \"\"\"Determine system scale based on requirements.\"\"\"\n if req.document_count \u003c 1_000_000 and req.queries_per_day \u003c 1_000:\n return Scale.SMALL\n elif req.document_count \u003c 100_000_000 and req.queries_per_day \u003c 100_000:\n return Scale.MEDIUM\n else:\n return Scale.LARGE\n \n def _recommend_chunking(self, req: Requirements, scale: Scale) -> ComponentRecommendation:\n \"\"\"Recommend chunking strategy.\"\"\"\n doc_types = set(req.document_types)\n \n if \"code\" in doc_types:\n strategy = \"semantic_code_aware\"\n config = {\"max_size\": 1000, \"preserve_functions\": True, \"overlap\": 50}\n rationale = \"Code documents benefit from function/class boundary awareness\"\n elif \"technical\" in doc_types or \"scientific\" in doc_types:\n strategy = \"semantic_heading_aware\" \n config = {\"max_size\": 1500, \"heading_weight\": 2.0, \"overlap\": 100}\n rationale = \"Technical documents have clear hierarchical structure\"\n elif len(doc_types) > 2 or \"mixed\" in doc_types:\n strategy = \"adaptive_chunking\"\n config = {\"strategies\": [\"paragraph\", \"sentence\", \"fixed\"], \"auto_select\": True}\n rationale = \"Mixed document types require adaptive strategy selection\"\n else:\n if req.avg_document_size > 5000:\n strategy = \"paragraph_based\"\n config = {\"max_size\": 2000, \"min_paragraph_size\": 100}\n rationale = \"Large documents benefit from paragraph-based chunking\"\n else:\n strategy = \"sentence_based\" \n config = {\"max_size\": 1000, \"sentence_overlap\": 1}\n rationale = \"Small to medium documents work well with sentence chunking\"\n \n return ComponentRecommendation(\n name=strategy,\n type=\"chunking\",\n config=config,\n rationale=rationale,\n pros=self._get_chunking_pros(strategy),\n cons=self._get_chunking_cons(strategy),\n cost_monthly=0.0 # Processing cost only\n )\n \n def _recommend_embedding(self, req: Requirements, scale: Scale) -> ComponentRecommendation:\n \"\"\"Recommend embedding model.\"\"\"\n doc_types = set(req.document_types)\n \n # Consider accuracy vs cost priority\n high_accuracy = req.accuracy_priority > 0.7\n cost_sensitive = req.cost_priority > 0.6\n \n if \"code\" in doc_types:\n if high_accuracy and not cost_sensitive:\n model = \"openai-code-search-ada-002\"\n cost_per_1k_tokens = 0.0001\n dimensions = 1536\n else:\n model = \"sentence-transformers/code-bert-base\"\n cost_per_1k_tokens = 0.0 # Self-hosted\n dimensions = 768\n elif \"scientific\" in doc_types:\n if high_accuracy:\n model = \"openai-text-embedding-ada-002\"\n cost_per_1k_tokens = 0.0001\n dimensions = 1536\n else:\n model = \"sentence-transformers/scibert-nli\" \n cost_per_1k_tokens = 0.0\n dimensions = 768\n else:\n if cost_sensitive or scale == Scale.SMALL:\n model = \"sentence-transformers/all-MiniLM-L6-v2\"\n cost_per_1k_tokens = 0.0\n dimensions = 384\n elif high_accuracy:\n model = \"openai-text-embedding-ada-002\"\n cost_per_1k_tokens = 0.0001 \n dimensions = 1536\n else:\n model = \"sentence-transformers/all-mpnet-base-v2\"\n cost_per_1k_tokens = 0.0\n dimensions = 768\n \n # Calculate monthly embedding cost\n total_tokens = req.document_count * (req.avg_document_size / 4) # ~4 chars per token\n query_tokens = req.queries_per_day * 30 * 20 # ~20 tokens per query per month\n monthly_cost = (total_tokens + query_tokens) * cost_per_1k_tokens / 1000\n \n return ComponentRecommendation(\n name=model,\n type=\"embedding\",\n config={\n \"model\": model,\n \"dimensions\": dimensions,\n \"batch_size\": 100 if scale == Scale.SMALL else 1000,\n \"cache_embeddings\": True\n },\n rationale=f\"Selected for {doc_types} with accuracy priority {req.accuracy_priority}\",\n pros=self._get_embedding_pros(model),\n cons=self._get_embedding_cons(model),\n cost_monthly=monthly_cost\n )\n \n def _recommend_vector_db(self, req: Requirements, scale: Scale) -> ComponentRecommendation:\n \"\"\"Recommend vector database.\"\"\"\n if scale == Scale.SMALL and req.cost_priority > 0.7:\n db = \"chroma\"\n cost = 0.0\n rationale = \"Local/embedded database suitable for small scale and cost optimization\"\n elif scale == Scale.SMALL and req.maintenance_complexity == \"low\":\n db = \"pgvector\"\n cost = 50.0 # PostgreSQL hosting\n rationale = \"Leverage existing PostgreSQL infrastructure\"\n elif scale == Scale.LARGE or req.latency_requirement == \"real_time\":\n db = \"pinecone\"\n vectors = req.document_count * 2 # Account for chunking\n cost = max(70, vectors * 0.00005) # $70 base + $0.00005 per vector\n rationale = \"Managed service with excellent performance for large scale\"\n elif req.maintenance_complexity == \"low\":\n db = \"weaviate_cloud\"\n vectors = req.document_count * 2\n cost = max(25, vectors * 0.00003)\n rationale = \"Managed Weaviate with good balance of features and cost\"\n else:\n db = \"qdrant\"\n cost = 100.0 # Self-hosted infrastructure estimate\n rationale = \"High performance self-hosted option with good scaling\"\n \n return ComponentRecommendation(\n name=db,\n type=\"vector_database\",\n config=self._get_vector_db_config(db, req, scale),\n rationale=rationale,\n pros=self._get_vector_db_pros(db),\n cons=self._get_vector_db_cons(db),\n cost_monthly=cost\n )\n \n def _recommend_retrieval(self, req: Requirements, scale: Scale) -> ComponentRecommendation:\n \"\"\"Recommend retrieval strategy.\"\"\"\n if req.accuracy_priority > 0.8:\n strategy = \"hybrid\"\n rationale = \"Hybrid retrieval for maximum accuracy combining dense and sparse methods\"\n elif \"technical\" in req.document_types or \"code\" in req.document_types:\n strategy = \"hybrid\"\n rationale = \"Technical content benefits from both semantic and keyword matching\"\n elif req.latency_requirement == \"real_time\":\n strategy = \"dense\"\n rationale = \"Dense retrieval faster for real-time requirements\"\n else:\n strategy = \"dense\"\n rationale = \"Dense retrieval suitable for general text search\"\n \n return ComponentRecommendation(\n name=strategy,\n type=\"retrieval\", \n config={\n \"strategy\": strategy,\n \"dense_weight\": 0.7 if strategy == \"hybrid\" else 1.0,\n \"sparse_weight\": 0.3 if strategy == \"hybrid\" else 0.0,\n \"top_k\": 20 if req.accuracy_priority > 0.7 else 10,\n \"similarity_threshold\": 0.7\n },\n rationale=rationale,\n pros=self._get_retrieval_pros(strategy),\n cons=self._get_retrieval_cons(strategy),\n cost_monthly=0.0\n )\n \n def _recommend_reranking(self, req: Requirements, scale: Scale) -> Optional[ComponentRecommendation]:\n \"\"\"Recommend reranking if beneficial.\"\"\"\n if req.accuracy_priority \u003c 0.6 or req.latency_requirement == \"real_time\":\n return None\n \n if req.cost_priority > 0.8:\n return None\n \n # Estimate reranking queries per month\n monthly_queries = req.queries_per_day * 30\n cost_per_query = 0.002 # Estimated cost for cross-encoder reranking\n monthly_cost = monthly_queries * cost_per_query\n \n if monthly_cost > req.budget_monthly * 0.3: # Don't exceed 30% of budget\n return None\n \n return ComponentRecommendation(\n name=\"cross_encoder_reranking\",\n type=\"reranking\",\n config={\n \"model\": \"cross-encoder/ms-marco-MiniLM-L-12-v2\",\n \"rerank_top_k\": 20,\n \"return_top_k\": 5,\n \"batch_size\": 16\n },\n rationale=\"Reranking improves precision for high-accuracy requirements\",\n pros=[\"Higher precision\", \"Better ranking quality\", \"Handles complex queries\"],\n cons=[\"Additional latency\", \"Higher cost\", \"More complexity\"],\n cost_monthly=monthly_cost\n )\n \n def _recommend_evaluation(self, req: Requirements, scale: Scale) -> ComponentRecommendation:\n \"\"\"Recommend evaluation framework.\"\"\"\n return ComponentRecommendation(\n name=\"comprehensive_evaluation\",\n type=\"evaluation\",\n config={\n \"metrics\": [\"precision@k\", \"recall@k\", \"mrr\", \"ndcg\"],\n \"k_values\": [1, 3, 5, 10],\n \"faithfulness_check\": True,\n \"relevance_scoring\": True,\n \"evaluation_frequency\": \"weekly\" if scale == Scale.LARGE else \"monthly\",\n \"sample_size\": min(1000, req.queries_per_day * 7)\n },\n rationale=\"Comprehensive evaluation essential for production RAG systems\",\n pros=[\"Quality monitoring\", \"Performance tracking\", \"Issue detection\"],\n cons=[\"Additional overhead\", \"Requires ground truth data\"],\n cost_monthly=20.0 # Evaluation tooling and compute\n )\n \n def _generate_architecture_diagram(self, chunking: ComponentRecommendation, \n embedding: ComponentRecommendation,\n vector_db: ComponentRecommendation,\n retrieval: ComponentRecommendation,\n reranking: Optional[ComponentRecommendation],\n evaluation: ComponentRecommendation) -> str:\n \"\"\"Generate Mermaid architecture diagram.\"\"\"\n \n diagram = \"\"\"```mermaid\ngraph TB\n %% Document Processing Pipeline\n A[Document Corpus] --> B[Document Chunking]\n B --> C[Embedding Generation]\n C --> D[Vector Database Storage]\n \n %% Query Processing Pipeline \n E[User Query] --> F[Query Processing]\n F --> G[Vector Search]\n D --> G\n G --> H[Retrieved Chunks]\n\"\"\"\n \n if reranking:\n diagram += \" H --> I[Reranking]\\n I --> J[Final Results]\\n\"\n else:\n diagram += \" H --> J[Final Results]\\n\"\n \n diagram += \"\"\" \n %% Evaluation Pipeline\n J --> K[Response Generation]\n K --> L[Evaluation Metrics]\n \n %% Component Details\n B -.-> B1[Strategy: \"\"\" + chunking.name + \"\"\"]\n C -.-> C1[Model: \"\"\" + embedding.name + \"\"\"]\n D -.-> D1[Database: \"\"\" + vector_db.name + \"\"\"]\n G -.-> G1[Method: \"\"\" + retrieval.name + \"\"\"]\n\"\"\"\n \n if reranking:\n diagram += \" I -.-> I1[Model: \" + reranking.name + \"]\\n\"\n \n diagram += \" L -.-> L1[Framework: \" + evaluation.name + \"]\\n```\"\n \n return diagram\n \n def _generate_config_templates(self, *components) -> Dict[str, Any]:\n \"\"\"Generate configuration templates for all components.\"\"\"\n configs = {}\n \n for component in components:\n if component:\n configs[component.type] = {\n \"component\": component.name,\n \"config\": component.config,\n \"rationale\": component.rationale\n }\n \n # Add deployment configuration\n configs[\"deployment\"] = {\n \"infrastructure\": \"cloud\" if any(\"pinecone\" in str(c.name) for c in components if c) else \"hybrid\",\n \"scaling\": {\n \"auto_scaling\": True,\n \"min_replicas\": 1,\n \"max_replicas\": 10\n },\n \"monitoring\": {\n \"metrics\": [\"latency\", \"throughput\", \"accuracy\"],\n \"alerts\": [\"high_latency\", \"low_accuracy\", \"service_down\"]\n }\n }\n \n return configs\n \n def _load_embedding_models(self) -> Dict[str, Dict[str, Any]]:\n \"\"\"Load embedding model specifications.\"\"\"\n return {\n \"openai-text-embedding-ada-002\": {\n \"dimensions\": 1536,\n \"cost_per_1k_tokens\": 0.0001,\n \"quality\": \"high\",\n \"speed\": \"medium\"\n },\n \"sentence-transformers/all-mpnet-base-v2\": {\n \"dimensions\": 768, \n \"cost_per_1k_tokens\": 0.0,\n \"quality\": \"high\",\n \"speed\": \"medium\"\n },\n \"sentence-transformers/all-MiniLM-L6-v2\": {\n \"dimensions\": 384,\n \"cost_per_1k_tokens\": 0.0,\n \"quality\": \"medium\",\n \"speed\": \"fast\"\n }\n }\n \n def _load_vector_databases(self) -> Dict[str, Dict[str, Any]]:\n \"\"\"Load vector database specifications.\"\"\"\n return {\n \"pinecone\": {\"managed\": True, \"scaling\": \"excellent\", \"cost\": \"high\"},\n \"weaviate\": {\"managed\": False, \"scaling\": \"good\", \"cost\": \"medium\"},\n \"qdrant\": {\"managed\": False, \"scaling\": \"excellent\", \"cost\": \"low\"},\n \"chroma\": {\"managed\": False, \"scaling\": \"poor\", \"cost\": \"free\"},\n \"pgvector\": {\"managed\": False, \"scaling\": \"good\", \"cost\": \"medium\"}\n }\n \n def _load_chunking_strategies(self) -> Dict[str, Dict[str, Any]]:\n \"\"\"Load chunking strategy specifications.\"\"\"\n return {\n \"fixed_size\": {\"complexity\": \"low\", \"quality\": \"medium\"},\n \"sentence_based\": {\"complexity\": \"medium\", \"quality\": \"good\"}, \n \"paragraph_based\": {\"complexity\": \"medium\", \"quality\": \"good\"},\n \"semantic_heading_aware\": {\"complexity\": \"high\", \"quality\": \"excellent\"}\n }\n \n def _get_vector_db_config(self, db: str, req: Requirements, scale: Scale) -> Dict[str, Any]:\n \"\"\"Get vector database configuration.\"\"\"\n base_config = {\n \"collection_name\": \"rag_documents\",\n \"distance_metric\": \"cosine\",\n \"index_type\": \"hnsw\"\n }\n \n if db == \"pinecone\":\n base_config.update({\n \"environment\": \"us-east1-gcp\",\n \"replicas\": 1 if scale == Scale.SMALL else 2,\n \"shards\": 1 if scale != Scale.LARGE else 3\n })\n elif db == \"qdrant\":\n base_config.update({\n \"memory_mapping\": True,\n \"quantization\": scale == Scale.LARGE,\n \"replication_factor\": 1 if scale == Scale.SMALL else 2\n })\n \n return base_config\n \n def _get_chunking_pros(self, strategy: str) -> List[str]:\n \"\"\"Get pros for chunking strategy.\"\"\"\n pros_map = {\n \"semantic_heading_aware\": [\"Preserves document structure\", \"High semantic coherence\", \"Good for technical docs\"],\n \"paragraph_based\": [\"Respects natural boundaries\", \"Good balance\", \"Readable chunks\"],\n \"sentence_based\": [\"Natural language boundaries\", \"Consistent quality\", \"Good for general text\"],\n \"fixed_size\": [\"Predictable sizes\", \"Simple implementation\", \"Consistent processing\"],\n \"adaptive_chunking\": [\"Handles mixed content\", \"Optimizes per document\", \"Best quality\"]\n }\n return pros_map.get(strategy, [\"Good general purpose strategy\"])\n \n def _get_chunking_cons(self, strategy: str) -> List[str]:\n \"\"\"Get cons for chunking strategy.\"\"\"\n cons_map = {\n \"semantic_heading_aware\": [\"Complex implementation\", \"May create large chunks\", \"Document-dependent\"],\n \"paragraph_based\": [\"Variable sizes\", \"May break context\", \"Document-dependent\"],\n \"sentence_based\": [\"May create small chunks\", \"Sentence detection issues\", \"Variable sizes\"],\n \"fixed_size\": [\"Breaks semantic boundaries\", \"May split sentences\", \"Context loss\"],\n \"adaptive_chunking\": [\"High complexity\", \"Slower processing\", \"Harder to debug\"]\n }\n return cons_map.get(strategy, [\"May not fit all use cases\"])\n \n def _get_embedding_pros(self, model: str) -> List[str]:\n \"\"\"Get pros for embedding model.\"\"\"\n if \"openai\" in model:\n return [\"High quality\", \"Regular updates\", \"Good performance\"]\n elif \"all-mpnet\" in model:\n return [\"High quality\", \"Free to use\", \"Good balance\"]\n elif \"MiniLM\" in model:\n return [\"Fast processing\", \"Small size\", \"Good for real-time\"]\n else:\n return [\"Specialized for domain\", \"Good performance\"]\n \n def _get_embedding_cons(self, model: str) -> List[str]:\n \"\"\"Get cons for embedding model.\"\"\"\n if \"openai\" in model:\n return [\"API costs\", \"Vendor lock-in\", \"Rate limits\"]\n elif \"sentence-transformers\" in model:\n return [\"Self-hosting required\", \"Model updates needed\", \"GPU beneficial\"]\n else:\n return [\"May require fine-tuning\", \"Domain-specific\"]\n \n def _get_vector_db_pros(self, db: str) -> List[str]:\n \"\"\"Get pros for vector database.\"\"\"\n pros_map = {\n \"pinecone\": [\"Fully managed\", \"Excellent performance\", \"Auto-scaling\"],\n \"weaviate\": [\"Rich features\", \"GraphQL API\", \"Multi-modal\"],\n \"qdrant\": [\"High performance\", \"Rust-based\", \"Good scaling\"],\n \"chroma\": [\"Simple setup\", \"Free\", \"Good for development\"],\n \"pgvector\": [\"SQL integration\", \"ACID compliance\", \"Familiar\"]\n }\n return pros_map.get(db, [\"Good performance\"])\n \n def _get_vector_db_cons(self, db: str) -> List[str]:\n \"\"\"Get cons for vector database.\"\"\"\n cons_map = {\n \"pinecone\": [\"Expensive\", \"Vendor lock-in\", \"Limited customization\"],\n \"weaviate\": [\"Complex setup\", \"Learning curve\", \"Resource intensive\"],\n \"qdrant\": [\"Self-managed\", \"Smaller community\", \"Setup complexity\"],\n \"chroma\": [\"Limited scaling\", \"Not production-ready\", \"Basic features\"],\n \"pgvector\": [\"PostgreSQL knowledge needed\", \"Less specialized\", \"Manual optimization\"]\n }\n return cons_map.get(db, [\"Requires maintenance\"])\n \n def _get_retrieval_pros(self, strategy: str) -> List[str]:\n \"\"\"Get pros for retrieval strategy.\"\"\"\n pros_map = {\n \"dense\": [\"Semantic understanding\", \"Good for paraphrases\", \"Fast\"],\n \"sparse\": [\"Exact matching\", \"Interpretable\", \"Good for keywords\"], \n \"hybrid\": [\"Best of both\", \"High accuracy\", \"Robust\"]\n }\n return pros_map.get(strategy, [\"Good performance\"])\n \n def _get_retrieval_cons(self, strategy: str) -> List[str]:\n \"\"\"Get cons for retrieval strategy.\"\"\"\n cons_map = {\n \"dense\": [\"May miss exact matches\", \"Embedding dependent\", \"Less interpretable\"],\n \"sparse\": [\"Vocabulary mismatch\", \"No semantic understanding\", \"Synonym issues\"],\n \"hybrid\": [\"More complex\", \"Tuning required\", \"Higher latency\"]\n }\n return cons_map.get(strategy, [\"May require tuning\"])\n\n\ndef load_requirements(file_path: str) -> Requirements:\n \"\"\"Load requirements from JSON file.\"\"\"\n with open(file_path, 'r') as f:\n data = json.load(f)\n \n return Requirements(**data)\n\n\ndef save_design(design: PipelineDesign, output_path: str):\n \"\"\"Save pipeline design to JSON file.\"\"\"\n # Convert to dict for JSON serialization\n design_dict = {}\n \n for field_name in design.__dataclass_fields__:\n value = getattr(design, field_name)\n if isinstance(value, ComponentRecommendation):\n design_dict[field_name] = asdict(value)\n elif value is None:\n design_dict[field_name] = None\n else:\n design_dict[field_name] = value\n \n with open(output_path, 'w') as f:\n json.dump(design_dict, f, indent=2)\n\n\ndef print_design_summary(design: PipelineDesign):\n \"\"\"Print human-readable design summary.\"\"\"\n print(\"\\n\" + \"=\"*60)\n print(\"RAG PIPELINE DESIGN SUMMARY\")\n print(\"=\"*60)\n \n print(f\"\\n💰 Total Monthly Cost: ${design.total_cost:.2f}\")\n \n print(f\"\\n🔧 Component Recommendations:\")\n components = [design.chunking, design.embedding, design.vector_db, \n design.retrieval, design.reranking, design.evaluation]\n \n for component in components:\n if component:\n print(f\"\\n {component.type.upper()}: {component.name}\")\n print(f\" Rationale: {component.rationale}\")\n if component.cost_monthly > 0:\n print(f\" Monthly Cost: ${component.cost_monthly:.2f}\")\n \n print(f\"\\n📊 Architecture Diagram:\")\n print(design.architecture_diagram)\n\n\ndef main():\n \"\"\"Main function with command-line interface.\"\"\"\n parser = argparse.ArgumentParser(description='Design RAG pipeline based on requirements')\n parser.add_argument('requirements', help='JSON file containing system requirements')\n parser.add_argument('--output', '-o', help='Output file for pipeline design (JSON)')\n parser.add_argument('--verbose', '-v', action='store_true', help='Verbose output')\n \n args = parser.parse_args()\n \n try:\n # Load requirements\n print(\"Loading requirements...\")\n requirements = load_requirements(args.requirements)\n \n # Design pipeline\n designer = RAGPipelineDesigner()\n design = designer.design_pipeline(requirements)\n \n # Save design\n if args.output:\n save_design(design, args.output)\n print(f\"Pipeline design saved to {args.output}\")\n \n # Print summary\n print_design_summary(design)\n \n if args.verbose:\n print(f\"\\n📋 Configuration Templates:\")\n for component_type, config in design.config_templates.items():\n print(f\"\\n {component_type.upper()}:\")\n print(f\" {json.dumps(config, indent=4)}\")\n \n except Exception as e:\n print(f\"Error: {e}\")\n return 1\n \n return 0\n\n\nif __name__ == '__main__':\n exit(main())","content_type":"text/x-python; charset=utf-8","language":"python","size":27852,"content_sha256":"c285db2e2642e4001d480b26809208f92f54e4c0a6dcf021f06d53b535a536f1"},{"filename":"references/chunking_strategies_comparison.md","content":"# Chunking Strategies Comparison\n\n## Executive Summary\n\nDocument chunking is the foundation of effective RAG systems. This analysis compares five primary chunking strategies across key metrics including semantic coherence, boundary quality, processing speed, and implementation complexity.\n\n## Strategies Analyzed\n\n### 1. Fixed-Size Chunking\n\n**Approach**: Split documents into chunks of predetermined size (characters/tokens) with optional overlap.\n\n**Variants**:\n- Character-based: 512, 1024, 2048 characters\n- Token-based: 128, 256, 512 tokens \n- Overlap: 0%, 10%, 20%\n\n**Performance Metrics**:\n- Processing Speed: ⭐⭐⭐⭐⭐ (Fastest)\n- Boundary Quality: ⭐⭐ (Poor - breaks mid-sentence)\n- Semantic Coherence: ⭐⭐ (Low - ignores content structure)\n- Implementation: ⭐⭐⭐⭐⭐ (Simplest)\n- Memory Efficiency: ⭐⭐⭐⭐⭐ (Predictable sizes)\n\n**Best For**: \n- Large-scale processing where speed is critical\n- Uniform document types\n- When consistent chunk sizes are required\n\n**Avoid When**:\n- Document quality varies significantly\n- Preserving context is critical\n- Processing narrative or technical content\n\n### 2. Sentence-Based Chunking\n\n**Approach**: Group complete sentences until size threshold reached, ensuring natural language boundaries.\n\n**Implementation Details**:\n- Sentence detection using regex patterns or NLP libraries\n- Size limits: 500-1500 characters typically\n- Overlap: 1-2 sentences for context preservation\n\n**Performance Metrics**:\n- Processing Speed: ⭐⭐⭐⭐ (Fast)\n- Boundary Quality: ⭐⭐⭐⭐ (Good - respects sentence boundaries)\n- Semantic Coherence: ⭐⭐⭐ (Medium - sentences may be topically unrelated)\n- Implementation: ⭐⭐⭐ (Moderate complexity)\n- Memory Efficiency: ⭐⭐⭐ (Variable sizes)\n\n**Best For**:\n- Narrative text (articles, books, blogs)\n- General-purpose text processing\n- When readability of chunks is important\n\n**Avoid When**:\n- Documents have complex sentence structures\n- Technical content with code/formulas\n- Very short or very long sentences dominate\n\n### 3. Paragraph-Based Chunking\n\n**Approach**: Use paragraph boundaries as primary split points, combining or splitting paragraphs based on size constraints.\n\n**Implementation Details**:\n- Paragraph detection via double newlines or HTML tags\n- Size limits: 1000-3000 characters\n- Hierarchical splitting for oversized paragraphs\n\n**Performance Metrics**:\n- Processing Speed: ⭐⭐⭐⭐ (Fast)\n- Boundary Quality: ⭐⭐⭐⭐⭐ (Excellent - natural breaks)\n- Semantic Coherence: ⭐⭐⭐⭐ (Good - paragraphs often topically coherent)\n- Implementation: ⭐⭐⭐ (Moderate complexity)\n- Memory Efficiency: ⭐⭐ (Highly variable sizes)\n\n**Best For**:\n- Well-structured documents\n- Articles and reports with clear paragraphs\n- When topic coherence is important\n\n**Avoid When**:\n- Documents have inconsistent paragraph structure\n- Paragraphs are extremely long or short\n- Technical documentation with mixed content\n\n### 4. Semantic Chunking (Heading-Aware)\n\n**Approach**: Use document structure (headings, sections) and semantic similarity to create topically coherent chunks.\n\n**Implementation Details**:\n- Heading detection (markdown, HTML, or inferred)\n- Topic modeling for section boundaries\n- Recursive splitting respecting hierarchy\n\n**Performance Metrics**:\n- Processing Speed: ⭐⭐ (Slow - requires analysis)\n- Boundary Quality: ⭐⭐⭐⭐⭐ (Excellent - respects document structure)\n- Semantic Coherence: ⭐⭐⭐⭐⭐ (Excellent - maintains topic coherence)\n- Implementation: ⭐⭐ (Complex)\n- Memory Efficiency: ⭐⭐ (Highly variable)\n\n**Best For**:\n- Technical documentation\n- Academic papers\n- Structured reports\n- When document hierarchy is important\n\n**Avoid When**:\n- Documents lack clear structure\n- Processing speed is critical\n- Implementation complexity must be minimized\n\n### 5. Recursive Chunking\n\n**Approach**: Hierarchical splitting using multiple strategies, preferring larger chunks when possible.\n\n**Implementation Details**:\n- Try larger chunks first (sections, paragraphs)\n- Recursively split if size exceeds threshold\n- Fallback hierarchy: document → section → paragraph → sentence → character\n\n**Performance Metrics**:\n- Processing Speed: ⭐⭐ (Slow - multiple passes)\n- Boundary Quality: ⭐⭐⭐⭐ (Good - adapts to content)\n- Semantic Coherence: ⭐⭐⭐⭐ (Good - preserves context when possible)\n- Implementation: ⭐⭐ (Complex logic)\n- Memory Efficiency: ⭐⭐⭐ (Optimizes chunk count)\n\n**Best For**:\n- Mixed document types\n- When chunk count optimization is important\n- Complex document structures\n\n**Avoid When**:\n- Simple, uniform documents\n- Real-time processing requirements\n- Debugging and maintenance overhead is a concern\n\n## Comparative Analysis\n\n### Chunk Size Distribution\n\n| Strategy | Mean Size | Std Dev | Min Size | Max Size | Coefficient of Variation |\n|----------|-----------|---------|----------|----------|-------------------------|\n| Fixed-Size | 1000 | 0 | 1000 | 1000 | 0.00 |\n| Sentence | 850 | 320 | 180 | 1500 | 0.38 |\n| Paragraph | 1200 | 680 | 200 | 3500 | 0.57 |\n| Semantic | 1400 | 920 | 300 | 4200 | 0.66 |\n| Recursive | 1100 | 450 | 400 | 2000 | 0.41 |\n\n### Processing Performance\n\n| Strategy | Processing Speed (docs/sec) | Memory Usage (MB/1K docs) | CPU Usage (%) |\n|----------|------------------------------|---------------------------|---------------|\n| Fixed-Size | 2500 | 50 | 15 |\n| Sentence | 1800 | 65 | 25 |\n| Paragraph | 2000 | 60 | 20 |\n| Semantic | 400 | 120 | 60 |\n| Recursive | 600 | 100 | 45 |\n\n### Quality Metrics\n\n| Strategy | Boundary Quality | Semantic Coherence | Context Preservation |\n|----------|------------------|-------------------|---------------------|\n| Fixed-Size | 0.15 | 0.32 | 0.28 |\n| Sentence | 0.85 | 0.58 | 0.65 |\n| Paragraph | 0.92 | 0.75 | 0.78 |\n| Semantic | 0.95 | 0.88 | 0.85 |\n| Recursive | 0.88 | 0.82 | 0.80 |\n\n## Domain-Specific Recommendations\n\n### Technical Documentation\n**Primary**: Semantic (heading-aware)\n**Secondary**: Recursive\n**Rationale**: Technical docs have clear hierarchical structure that should be preserved\n\n### Scientific Papers \n**Primary**: Semantic (heading-aware)\n**Secondary**: Paragraph-based\n**Rationale**: Papers have sections (abstract, methodology, results) that form coherent units\n\n### News Articles\n**Primary**: Paragraph-based\n**Secondary**: Sentence-based\n**Rationale**: Inverted pyramid structure means paragraphs are typically topically coherent\n\n### Legal Documents\n**Primary**: Paragraph-based\n**Secondary**: Semantic\n**Rationale**: Legal text has specific paragraph structures that shouldn't be broken\n\n### Code Documentation\n**Primary**: Semantic (code-aware)\n**Secondary**: Recursive\n**Rationale**: Code blocks, functions, and classes form natural boundaries\n\n### General Web Content\n**Primary**: Sentence-based\n**Secondary**: Paragraph-based\n**Rationale**: Variable quality and structure require robust general-purpose approach\n\n## Implementation Guidelines\n\n### Choosing Chunk Size\n\n1. **Consider retrieval context**: Smaller chunks (500-800 chars) for precise retrieval\n2. **Consider generation context**: Larger chunks (1000-2000 chars) for comprehensive answers\n3. **Model context limits**: Ensure chunks fit in embedding model context window\n4. **Query patterns**: Specific queries need smaller chunks, broad queries benefit from larger\n\n### Overlap Configuration\n\n- **None (0%)**: When context bleeding is problematic\n- **Low (5-10%)**: General-purpose overlap for context continuity\n- **Medium (15-20%)**: When context preservation is critical\n- **High (25%+)**: Rarely beneficial, increases storage costs significantly\n\n### Metadata Preservation\n\nAlways preserve:\n- Document source/path\n- Chunk position/sequence\n- Heading hierarchy (if applicable)\n- Creation/modification timestamps\n\nConditionally preserve:\n- Page numbers (for PDFs)\n- Section titles\n- Author information\n- Document type/category\n\n## Evaluation Framework\n\n### Automated Metrics\n\n1. **Chunk Size Consistency**: Standard deviation of chunk sizes\n2. **Boundary Quality Score**: Fraction of chunks ending with complete sentences\n3. **Topic Coherence**: Average cosine similarity between consecutive chunks\n4. **Processing Speed**: Documents processed per second\n5. **Memory Efficiency**: Peak memory usage during processing\n\n### Manual Evaluation\n\n1. **Readability**: Can humans easily understand chunk content?\n2. **Completeness**: Do chunks contain complete thoughts/concepts?\n3. **Context Sufficiency**: Is enough context preserved for accurate retrieval?\n4. **Boundary Appropriateness**: Do chunk boundaries make semantic sense?\n\n### A/B Testing Framework\n\n1. **Baseline Setup**: Establish current chunking strategy performance\n2. **Metric Selection**: Choose relevant metrics (precision@k, user satisfaction)\n3. **Sample Size**: Ensure statistical significance (typically 1000+ queries)\n4. **Duration**: Run for sufficient time to capture usage patterns\n5. **Analysis**: Statistical significance testing and practical effect size\n\n## Cost-Benefit Analysis\n\n### Development Costs\n- Fixed-Size: 1 developer-day\n- Sentence-Based: 3-5 developer-days\n- Paragraph-Based: 3-5 developer-days\n- Semantic: 10-15 developer-days\n- Recursive: 15-20 developer-days\n\n### Operational Costs\n- Processing overhead: Semantic chunking 3-5x slower than fixed-size\n- Storage overhead: Variable-size chunks may waste storage slots\n- Maintenance overhead: Complex strategies require more monitoring\n\n### Quality Benefits\n- Retrieval accuracy improvement: 10-30% for semantic vs fixed-size\n- User satisfaction: Measurable improvement with better chunk boundaries\n- Downstream task performance: Better chunks improve generation quality\n\n## Conclusion\n\nThe optimal chunking strategy depends on your specific use case:\n\n- **Speed-critical systems**: Fixed-size chunking\n- **General-purpose applications**: Sentence-based chunking\n- **High-quality requirements**: Semantic or recursive chunking\n- **Mixed environments**: Adaptive strategy selection\n\nConsider implementing multiple strategies and A/B testing to determine the best approach for your specific document corpus and user queries.","content_type":"text/markdown; charset=utf-8","language":"markdown","size":10181,"content_sha256":"9660ca53d86521338e5e168a4fae83bf553c51e33d8f556cabb6f1ed12701b13"},{"filename":"references/embedding_model_benchmark.md","content":"# Embedding Model Benchmark 2024\n\n## Executive Summary\n\nThis comprehensive benchmark evaluates 15 popular embedding models across multiple dimensions including retrieval quality, processing speed, memory usage, and cost. Results are based on evaluation across 5 diverse datasets totaling 2M+ documents and 50K queries.\n\n## Models Evaluated\n\n### OpenAI Models\n- **text-embedding-ada-002** (1536 dim) - Latest general-purpose model\n- **text-embedding-3-small** (1536 dim) - Optimized for speed/cost\n- **text-embedding-3-large** (3072 dim) - Maximum quality\n\n### Sentence Transformers (Open Source)\n- **all-mpnet-base-v2** (768 dim) - High-quality general purpose\n- **all-MiniLM-L6-v2** (384 dim) - Fast and compact\n- **all-MiniLM-L12-v2** (384 dim) - Better quality than L6\n- **paraphrase-multilingual-mpnet-base-v2** (768 dim) - Multilingual\n- **multi-qa-mpnet-base-dot-v1** (768 dim) - Optimized for Q&A\n\n### Specialized Models\n- **sentence-transformers/msmarco-distilbert-base-v4** (768 dim) - Search-optimized\n- **intfloat/e5-large-v2** (1024 dim) - State-of-the-art open source\n- **BAAI/bge-large-en-v1.5** (1024 dim) - Chinese team, excellent performance\n- **thenlper/gte-large** (1024 dim) - Recent high-performer\n\n### Domain-Specific Models\n- **microsoft/codebert-base** (768 dim) - Code embeddings\n- **allenai/scibert_scivocab_uncased** (768 dim) - Scientific text\n- **microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract** (768 dim) - Biomedical\n\n## Evaluation Methodology\n\n### Datasets Used\n\n1. **MS MARCO Passage Ranking** (8.8M passages, 6,980 queries)\n - General web search scenarios\n - Factual and informational queries\n\n2. **Natural Questions** (307K passages, 3,452 queries) \n - Wikipedia-based question answering\n - Natural language queries\n\n3. **TREC-COVID** (171K scientific papers, 50 queries)\n - Biomedical/scientific literature search\n - Technical domain knowledge\n\n4. **FiQA-2018** (57K forum posts, 648 queries)\n - Financial domain question answering\n - Domain-specific terminology\n\n5. **ArguAna** (8.67K arguments, 1,406 queries)\n - Counter-argument retrieval\n - Reasoning and argumentation\n\n### Metrics Calculated\n\n- **Retrieval Quality**: NDCG@10, MRR@10, Recall@100\n- **Speed**: Queries per second, documents per second (encoding)\n- **Memory**: Peak RAM usage, model size on disk\n- **Cost**: API costs (for commercial models) or compute costs (for self-hosted)\n\n### Hardware Setup\n- **CPU**: Intel Xeon Gold 6248 (40 cores)\n- **GPU**: NVIDIA V100 32GB (for transformer models)\n- **RAM**: 256GB DDR4\n- **Storage**: NVMe SSD\n\n## Results Overview\n\n### Retrieval Quality Rankings\n\n| Rank | Model | NDCG@10 | MRR@10 | Recall@100 | Overall Score |\n|------|-------|---------|--------|------------|---------------|\n| 1 | text-embedding-3-large | 0.594 | 0.431 | 0.892 | 0.639 |\n| 2 | BAAI/bge-large-en-v1.5 | 0.588 | 0.425 | 0.885 | 0.633 |\n| 3 | intfloat/e5-large-v2 | 0.582 | 0.419 | 0.878 | 0.626 |\n| 4 | text-embedding-ada-002 | 0.578 | 0.415 | 0.871 | 0.621 |\n| 5 | thenlper/gte-large | 0.571 | 0.408 | 0.865 | 0.615 |\n| 6 | all-mpnet-base-v2 | 0.543 | 0.385 | 0.824 | 0.584 |\n| 7 | multi-qa-mpnet-base-dot-v1 | 0.538 | 0.381 | 0.818 | 0.579 |\n| 8 | text-embedding-3-small | 0.535 | 0.378 | 0.815 | 0.576 |\n| 9 | msmarco-distilbert-base-v4 | 0.529 | 0.372 | 0.805 | 0.569 |\n| 10 | all-MiniLM-L12-v2 | 0.498 | 0.348 | 0.765 | 0.537 |\n| 11 | all-MiniLM-L6-v2 | 0.476 | 0.331 | 0.738 | 0.515 |\n| 12 | paraphrase-multilingual-mpnet | 0.465 | 0.324 | 0.729 | 0.506 |\n\n### Speed Performance\n\n| Model | Encoding Speed (docs/sec) | Query Speed (queries/sec) | Latency (ms) |\n|-------|---------------------------|---------------------------|--------------|\n| all-MiniLM-L6-v2 | 14,200 | 2,850 | 0.35 |\n| all-MiniLM-L12-v2 | 8,950 | 1,790 | 0.56 |\n| text-embedding-3-small | 8,500* | 1,700* | 0.59* |\n| msmarco-distilbert-base-v4 | 6,800 | 1,360 | 0.74 |\n| all-mpnet-base-v2 | 2,840 | 568 | 1.76 |\n| multi-qa-mpnet-base-dot-v1 | 2,760 | 552 | 1.81 |\n| text-embedding-ada-002 | 2,500* | 500* | 2.00* |\n| paraphrase-multilingual-mpnet | 2,650 | 530 | 1.89 |\n| thenlper/gte-large | 1,420 | 284 | 3.52 |\n| intfloat/e5-large-v2 | 1,380 | 276 | 3.62 |\n| BAAI/bge-large-en-v1.5 | 1,350 | 270 | 3.70 |\n| text-embedding-3-large | 1,200* | 240* | 4.17* |\n\n*API-based models - speeds include network latency\n\n### Memory Usage\n\n| Model | Model Size (MB) | Peak RAM (GB) | GPU VRAM (GB) |\n|-------|-----------------|---------------|---------------|\n| all-MiniLM-L6-v2 | 91 | 1.2 | 2.1 |\n| all-MiniLM-L12-v2 | 134 | 1.8 | 3.2 |\n| msmarco-distilbert-base-v4 | 268 | 2.4 | 4.8 |\n| all-mpnet-base-v2 | 438 | 3.2 | 6.4 |\n| multi-qa-mpnet-base-dot-v1 | 438 | 3.2 | 6.4 |\n| paraphrase-multilingual-mpnet | 438 | 3.2 | 6.4 |\n| thenlper/gte-large | 670 | 4.8 | 8.6 |\n| intfloat/e5-large-v2 | 670 | 4.8 | 8.6 |\n| BAAI/bge-large-en-v1.5 | 670 | 4.8 | 8.6 |\n| OpenAI Models | N/A | 0.1 | 0.0 |\n\n### Cost Analysis (1M tokens processed)\n\n| Model | Type | Cost per 1M tokens | Monthly Cost (10M tokens) |\n|-------|------|--------------------|---------------------------|\n| text-embedding-3-small | API | $0.02 | $0.20 |\n| text-embedding-ada-002 | API | $0.10 | $1.00 |\n| text-embedding-3-large | API | $1.30 | $13.00 |\n| all-MiniLM-L6-v2 | Self-hosted | $0.05 | $0.50 |\n| all-MiniLM-L12-v2 | Self-hosted | $0.08 | $0.80 |\n| all-mpnet-base-v2 | Self-hosted | $0.15 | $1.50 |\n| intfloat/e5-large-v2 | Self-hosted | $0.25 | $2.50 |\n| BAAI/bge-large-en-v1.5 | Self-hosted | $0.25 | $2.50 |\n| thenlper/gte-large | Self-hosted | $0.25 | $2.50 |\n\n*Self-hosted costs include compute, not including initial setup\n\n## Detailed Analysis\n\n### Quality vs Speed Trade-offs\n\n**High Performance Tier** (NDCG@10 > 0.57):\n- text-embedding-3-large: Best quality, expensive, slow\n- BAAI/bge-large-en-v1.5: Excellent quality, free, moderate speed\n- intfloat/e5-large-v2: Great quality, free, moderate speed\n\n**Balanced Tier** (NDCG@10 = 0.54-0.57):\n- all-mpnet-base-v2: Good quality-speed balance, widely adopted\n- text-embedding-ada-002: Good quality, reasonable API cost\n- multi-qa-mpnet-base-dot-v1: Q&A optimized, good for RAG\n\n**Speed Tier** (NDCG@10 = 0.47-0.54):\n- all-MiniLM-L12-v2: Best small model, good for real-time\n- all-MiniLM-L6-v2: Fastest processing, acceptable quality\n\n### Domain-Specific Performance\n\n#### Scientific/Technical Documents (TREC-COVID)\n1. **allenai/scibert**: 0.612 NDCG@10 (+15% vs general models)\n2. **text-embedding-3-large**: 0.589 NDCG@10\n3. **BAAI/bge-large-en-v1.5**: 0.581 NDCG@10\n\n#### Code Search (Custom CodeSearchNet evaluation)\n1. **microsoft/codebert-base**: 0.547 NDCG@10 (+22% vs general models)\n2. **text-embedding-ada-002**: 0.492 NDCG@10\n3. **all-mpnet-base-v2**: 0.478 NDCG@10\n\n#### Financial Domain (FiQA-2018)\n1. **text-embedding-3-large**: 0.573 NDCG@10\n2. **intfloat/e5-large-v2**: 0.567 NDCG@10\n3. **BAAI/bge-large-en-v1.5**: 0.561 NDCG@10\n\n### Multilingual Capabilities\n\nTested on translated versions of Natural Questions (Spanish, French, German):\n\n| Model | English NDCG@10 | Multilingual Avg | Degradation |\n|-------|-----------------|------------------|-------------|\n| paraphrase-multilingual-mpnet | 0.465 | 0.448 | 3.7% |\n| text-embedding-3-large | 0.594 | 0.521 | 12.3% |\n| text-embedding-ada-002 | 0.578 | 0.495 | 14.4% |\n| intfloat/e5-large-v2 | 0.582 | 0.483 | 17.0% |\n\n## Recommendations by Use Case\n\n### High-Volume Production Systems\n**Primary**: BAAI/bge-large-en-v1.5\n- Excellent quality (2nd best overall)\n- No API costs or rate limits\n- Reasonable resource requirements\n\n**Secondary**: intfloat/e5-large-v2\n- Very close quality to bge-large\n- Active development community\n- Good documentation\n\n### Cost-Sensitive Applications \n**Primary**: all-MiniLM-L6-v2\n- Lowest operational cost\n- Fastest processing\n- Acceptable quality for many use cases\n\n**Secondary**: text-embedding-3-small\n- Better quality than MiniLM\n- Competitive API pricing\n- No infrastructure overhead\n\n### Maximum Quality Requirements\n**Primary**: text-embedding-3-large\n- Best overall quality\n- Latest OpenAI technology\n- Worth the cost for critical applications\n\n**Secondary**: BAAI/bge-large-en-v1.5\n- Nearly equivalent quality\n- No ongoing API costs\n- Full control over deployment\n\n### Real-Time Applications (\u003c 100ms latency)\n**Primary**: all-MiniLM-L6-v2\n- Sub-millisecond inference\n- Small memory footprint\n- Easy to scale horizontally\n\n**Alternative**: text-embedding-3-small (if API latency acceptable)\n- Better quality than MiniLM\n- Reasonable API speed\n- No infrastructure management\n\n### Domain-Specific Applications\n\n**Scientific/Research**: \n1. Domain-specific model (SciBERT, BioBERT) if available\n2. text-embedding-3-large for general scientific content\n3. intfloat/e5-large-v2 as open-source alternative\n\n**Code/Technical**: \n1. microsoft/codebert-base for code search\n2. text-embedding-ada-002 for mixed code/text\n3. all-mpnet-base-v2 for technical documentation\n\n**Multilingual**:\n1. paraphrase-multilingual-mpnet-base-v2 for balanced multilingual\n2. text-embedding-3-large with translation pipeline\n3. Language-specific models when available\n\n## Implementation Guidelines\n\n### Model Selection Framework\n\n1. **Define Quality Requirements**\n - Minimum acceptable NDCG@10 threshold\n - Critical vs non-critical application\n - User tolerance for imperfect results\n\n2. **Assess Performance Requirements**\n - Expected queries per second\n - Latency requirements (real-time vs batch)\n - Concurrent user load\n\n3. **Evaluate Resource Constraints**\n - Available GPU memory\n - CPU capabilities\n - Network bandwidth (for API models)\n\n4. **Consider Operational Factors**\n - Team expertise with model deployment\n - Monitoring and maintenance capabilities\n - Vendor lock-in tolerance\n\n### Deployment Patterns\n\n**Single Model Deployment**:\n- Simplest approach\n- Choose one model for all use cases\n- Optimize infrastructure for that model\n\n**Tiered Deployment**:\n- Fast model for initial filtering (MiniLM)\n- High-quality model for reranking (bge-large)\n- Balance speed and quality\n\n**Domain-Specific Routing**:\n- Route queries to specialized models\n- Code queries → CodeBERT\n- Scientific queries → SciBERT\n- General queries → general model\n\n### A/B Testing Strategy\n\n1. **Baseline Establishment**\n - Current model performance metrics\n - User satisfaction baselines\n - System performance baselines\n\n2. **Gradual Rollout**\n - 5% traffic to new model initially\n - Monitor key metrics closely\n - Gradual increase if positive results\n\n3. **Key Metrics to Track**\n - Retrieval quality (NDCG, MRR)\n - User engagement (click-through rates)\n - System performance (latency, errors)\n - Cost metrics (API calls, compute usage)\n\n## Future Considerations\n\n### Emerging Trends\n\n1. **Instruction-Tuned Embeddings**: Models fine-tuned for specific instruction types\n2. **Multimodal Embeddings**: Text + image + audio embeddings\n3. **Extreme Efficiency**: Sub-100MB models with competitive quality\n4. **Dynamic Embeddings**: Context-aware embeddings that adapt to queries\n\n### Model Evolution Tracking\n\n**OpenAI**: Regular model updates, expect 2-3 new releases per year\n**Open Source**: Rapid innovation, new SOTA models every 3-6 months\n**Specialized Models**: Domain-specific models becoming more common\n\n### Performance Optimization\n\n1. **Quantization**: 8-bit and 4-bit quantization for memory efficiency\n2. **ONNX Optimization**: Convert models for faster inference\n3. **Model Distillation**: Create smaller, faster versions of large models\n4. **Batch Optimization**: Optimize for batch processing vs single queries\n\n## Conclusion\n\nThe embedding model landscape offers excellent options across all use cases:\n\n- **Quality Leaders**: text-embedding-3-large, bge-large-en-v1.5, e5-large-v2\n- **Speed Champions**: all-MiniLM-L6-v2, text-embedding-3-small\n- **Cost Optimized**: Open source models (bge, e5, mpnet series)\n- **Specialized**: Domain-specific models when available\n\nThe key is matching your specific requirements to the right model characteristics. Consider starting with BAAI/bge-large-en-v1.5 as a strong general-purpose choice, then optimize based on your specific needs and constraints.","content_type":"text/markdown; charset=utf-8","language":"markdown","size":12170,"content_sha256":"033e3c0b94b2751b1289933e01791dc11415eabf2465a0f3286e122f51a93e34"},{"filename":"references/rag_evaluation_framework.md","content":"# RAG Evaluation Framework\n\n## Overview\n\nEvaluating Retrieval-Augmented Generation (RAG) systems requires a comprehensive approach that measures both retrieval quality and generation performance. This framework provides methodologies, metrics, and tools for systematic RAG evaluation across different stages of the pipeline.\n\n## Evaluation Dimensions\n\n### 1. Retrieval Quality (Information Retrieval Metrics)\n\n**Precision@K**: Fraction of retrieved documents that are relevant\n- Formula: `Precision@K = Relevant Retrieved@K / K`\n- Use Case: Measuring result quality at different cutoff points\n- Target Values: >0.7 for K=1, >0.5 for K=5, >0.3 for K=10\n\n**Recall@K**: Fraction of relevant documents that are retrieved\n- Formula: `Recall@K = Relevant Retrieved@K / Total Relevant`\n- Use Case: Measuring coverage of relevant information\n- Target Values: >0.8 for K=10, >0.9 for K=20\n\n**Mean Reciprocal Rank (MRR)**: Average reciprocal rank of first relevant result\n- Formula: `MRR = (1/Q) × Σ(1/rank_i)` where rank_i is position of first relevant result\n- Use Case: Measuring how quickly users find relevant information\n- Target Values: >0.6 for good systems, >0.8 for excellent systems\n\n**Normalized Discounted Cumulative Gain (NDCG@K)**: Position-aware relevance metric\n- Formula: `NDCG@K = DCG@K / IDCG@K`\n- Use Case: Penalizing relevant documents that appear lower in rankings\n- Target Values: >0.7 for K=5, >0.6 for K=10\n\n### 2. Generation Quality (RAG-Specific Metrics)\n\n**Faithfulness**: How well the generated answer is grounded in retrieved context\n- Measurement: NLI-based entailment scoring, fact verification\n- Implementation: Check if each claim in answer is supported by context\n- Target Values: >0.95 for factual systems, >0.85 for general applications\n\n**Answer Relevance**: How well the generated answer addresses the original question\n- Measurement: Semantic similarity between question and answer\n- Implementation: Embedding similarity, keyword overlap, LLM-as-judge\n- Target Values: >0.8 for focused answers, >0.7 for comprehensive responses\n\n**Context Relevance**: How relevant the retrieved context is to the question\n- Measurement: Relevance scoring of each retrieved chunk\n- Implementation: Question-context similarity, manual annotation\n- Target Values: >0.7 for average relevance of top-5 chunks\n\n**Context Precision**: Fraction of relevant sentences in retrieved context\n- Measurement: Sentence-level relevance annotation\n- Implementation: Binary classification of each sentence's relevance\n- Target Values: >0.6 for efficient context usage\n\n**Context Recall**: Coverage of necessary information for answering the question\n- Measurement: Whether all required facts are present in context\n- Implementation: Expert annotation or automated fact extraction\n- Target Values: >0.8 for comprehensive coverage\n\n### 3. End-to-End Quality\n\n**Correctness**: Factual accuracy of the generated answer\n- Measurement: Expert evaluation, automated fact-checking\n- Implementation: Compare against ground truth, verify claims\n- Scoring: Binary (correct/incorrect) or scaled (1-5)\n\n**Completeness**: Whether the answer addresses all aspects of the question\n- Measurement: Coverage of question components\n- Implementation: Aspect-based evaluation, expert annotation\n- Scoring: Fraction of question aspects covered\n\n**Helpfulness**: Overall utility of the response to the user\n- Measurement: User ratings, task completion rates\n- Implementation: Human evaluation, A/B testing\n- Scoring: 1-5 Likert scale or thumbs up/down\n\n## Evaluation Methodologies\n\n### 1. Offline Evaluation\n\n**Dataset Requirements**:\n- Diverse query set (100+ queries for statistical significance)\n- Ground truth relevance judgments\n- Reference answers (for generation evaluation)\n- Representative document corpus\n\n**Evaluation Pipeline**:\n1. Query Processing: Standardize query format and preprocessing\n2. Retrieval Execution: Run retrieval with consistent parameters\n3. Generation Execution: Generate answers using retrieved context\n4. Metric Calculation: Compute all relevant metrics\n5. Statistical Analysis: Significance testing, confidence intervals\n\n**Best Practices**:\n- Stratify queries by type (factual, analytical, conversational)\n- Include edge cases (ambiguous queries, no-answer situations)\n- Use multiple annotators with inter-rater agreement analysis\n- Regular re-evaluation as system evolves\n\n### 2. Online Evaluation (A/B Testing)\n\n**Metrics to Track**:\n- User engagement: Click-through rates, time on page\n- User satisfaction: Explicit ratings, implicit feedback\n- Task completion: Success rates for specific user goals\n- System performance: Latency, error rates\n\n**Experimental Design**:\n- Randomized assignment to treatment/control groups\n- Sufficient sample size (typically 1000+ users per group)\n- Runtime duration (1-4 weeks for stable results)\n- Proper randomization and bias mitigation\n\n### 3. Human Evaluation\n\n**Evaluation Aspects**:\n- Factual Accuracy: Is the information correct?\n- Relevance: Does the answer address the question?\n- Completeness: Are all aspects covered?\n- Clarity: Is the answer easy to understand?\n- Conciseness: Is the answer appropriately brief?\n\n**Annotation Guidelines**:\n- Clear scoring rubrics (e.g., 1-5 scales with examples)\n- Multiple annotators per sample (typically 3-5)\n- Training and calibration sessions\n- Regular quality checks and inter-rater agreement\n\n## Implementation Framework\n\n### 1. Automated Evaluation Pipeline\n\n```python\nclass RAGEvaluator:\n def __init__(self, retriever, generator, metrics_config):\n self.retriever = retriever\n self.generator = generator\n self.metrics = self._initialize_metrics(metrics_config)\n \n def evaluate_query(self, query, ground_truth):\n # Retrieval evaluation\n retrieved_docs = self.retriever.search(query)\n retrieval_metrics = self.evaluate_retrieval(\n retrieved_docs, ground_truth['relevant_docs']\n )\n \n # Generation evaluation\n generated_answer = self.generator.generate(query, retrieved_docs)\n generation_metrics = self.evaluate_generation(\n query, generated_answer, retrieved_docs, ground_truth['answer']\n )\n \n return {**retrieval_metrics, **generation_metrics}\n```\n\n### 2. Metric Implementations\n\n**Faithfulness Score**:\n```python\ndef calculate_faithfulness(answer, context):\n # Split answer into claims\n claims = extract_claims(answer)\n \n # Check each claim against context\n faithful_claims = 0\n for claim in claims:\n if is_supported_by_context(claim, context):\n faithful_claims += 1\n \n return faithful_claims / len(claims) if claims else 0\n```\n\n**Context Relevance Score**:\n```python\ndef calculate_context_relevance(query, contexts):\n relevance_scores = []\n for context in contexts:\n similarity = embedding_similarity(query, context)\n relevance_scores.append(similarity)\n \n return {\n 'average_relevance': mean(relevance_scores),\n 'top_k_relevance': mean(relevance_scores[:k]),\n 'relevance_distribution': relevance_scores\n }\n```\n\n### 3. Evaluation Dataset Creation\n\n**Query Collection Strategies**:\n1. **User Log Analysis**: Extract real user queries from production systems\n2. **Expert Generation**: Domain experts create representative queries\n3. **Synthetic Generation**: LLM-generated queries based on document content\n4. **Community Sourcing**: Crowdsourced query collection\n\n**Ground Truth Creation**:\n1. **Document Relevance**: Expert annotation of relevant documents per query\n2. **Answer Creation**: Expert-written reference answers\n3. **Aspect Annotation**: Mark which aspects of complex questions are addressed\n4. **Quality Control**: Multiple annotators with disagreement resolution\n\n## Evaluation Datasets and Benchmarks\n\n### 1. General Domain Benchmarks\n\n**MS MARCO**: Large-scale reading comprehension dataset\n- 100K real user queries from Bing search\n- Passage-level and document-level evaluation\n- Both retrieval and generation evaluation supported\n\n**Natural Questions**: Google search queries with Wikipedia answers\n- 307K training examples, 8K development examples\n- Natural language questions from real users\n- Both short and long answer evaluation\n\n**SQUAD 2.0**: Reading comprehension with unanswerable questions\n- 150K question-answer pairs\n- Includes questions that cannot be answered from context\n- Tests system's ability to recognize unanswerable queries\n\n### 2. Domain-Specific Benchmarks\n\n**TREC-COVID**: Scientific literature search\n- 50 queries on COVID-19 research topics\n- 171K scientific papers as corpus\n- Expert relevance judgments\n\n**FiQA**: Financial question answering\n- 648 questions from financial forums\n- 57K financial forum posts as corpus\n- Domain-specific terminology and concepts\n\n**BioASQ**: Biomedical semantic indexing and question answering\n- 3K biomedical questions\n- PubMed abstracts as corpus\n- Expert physician annotations\n\n### 3. Multilingual Benchmarks\n\n**Mr. TyDi**: Multilingual question answering\n- 11 languages including Arabic, Bengali, Korean\n- Wikipedia passages in each language\n- Cultural and linguistic diversity testing\n\n**MLQA**: Cross-lingual question answering\n- Questions in one language, answers in another\n- 7 languages with all pair combinations\n- Tests multilingual retrieval capabilities\n\n## Continuous Evaluation Framework\n\n### 1. Monitoring Pipeline\n\n**Real-time Metrics**:\n- System latency (p50, p95, p99)\n- Error rates and failure modes\n- User satisfaction scores\n- Query volume and patterns\n\n**Batch Evaluation**:\n- Weekly/monthly evaluation on test sets\n- Performance trend analysis\n- Regression detection\n- Model drift monitoring\n\n### 2. Quality Assurance\n\n**Automated Quality Checks**:\n- Hallucination detection\n- Toxicity and bias screening\n- Factual consistency verification\n- Output format validation\n\n**Human Review Process**:\n- Random sampling of responses (1-5% of production queries)\n- Expert review of edge cases and failures\n- User feedback integration\n- Regular calibration of automated metrics\n\n### 3. Performance Optimization\n\n**A/B Testing Framework**:\n- Infrastructure for controlled experiments\n- Statistical significance testing\n- Multi-armed bandit optimization\n- Gradual rollout procedures\n\n**Feedback Loop Integration**:\n- User feedback incorporation into training data\n- Error analysis and root cause identification\n- Iterative improvement processes\n- Model fine-tuning based on evaluation results\n\n## Tools and Libraries\n\n### 1. Open Source Tools\n\n**RAGAS**: RAG Assessment framework\n- Comprehensive metric implementations\n- Easy integration with popular RAG frameworks\n- Support for both synthetic and human evaluation\n\n**TruEra TruLens**: ML observability for RAG\n- Real-time monitoring and evaluation\n- Comprehensive metric tracking\n- Integration with popular vector databases\n\n**LangSmith**: LangChain evaluation and monitoring\n- End-to-end RAG pipeline evaluation\n- Human feedback integration\n- Performance analytics and debugging\n\n### 2. Commercial Solutions\n\n**Weights & Biases**: ML experiment tracking\n- A/B testing infrastructure\n- Comprehensive metrics dashboard\n- Team collaboration features\n\n**Neptune**: ML metadata store\n- Experiment comparison and analysis\n- Model performance monitoring\n- Integration with popular ML frameworks\n\n**Comet**: ML platform for tracking experiments\n- Real-time monitoring\n- Model comparison and selection\n- Automated report generation\n\n## Best Practices\n\n### 1. Evaluation Design\n\n**Metric Selection**:\n- Choose metrics aligned with business objectives\n- Use multiple complementary metrics\n- Include both automated and human evaluation\n- Consider computational cost vs. insight value\n\n**Dataset Preparation**:\n- Ensure representative query distribution\n- Include edge cases and failure modes\n- Maintain high annotation quality\n- Regular dataset updates and validation\n\n### 2. Statistical Rigor\n\n**Sample Sizes**:\n- Minimum 100 queries for basic evaluation\n- 1000+ queries for robust statistical analysis\n- Power analysis for A/B testing\n- Confidence interval reporting\n\n**Significance Testing**:\n- Use appropriate statistical tests (t-tests, Mann-Whitney U)\n- Multiple comparison corrections (Bonferroni, FDR)\n- Effect size reporting alongside p-values\n- Bootstrap confidence intervals for stability\n\n### 3. Operational Integration\n\n**Automated Pipelines**:\n- Continuous integration/deployment integration\n- Automated regression testing\n- Performance threshold enforcement\n- Alert systems for quality degradation\n\n**Human-in-the-Loop**:\n- Regular expert review processes\n- User feedback collection and analysis\n- Annotation quality control\n- Bias detection and mitigation\n\n## Common Pitfalls and Solutions\n\n### 1. Evaluation Bias\n\n**Problem**: Test set not representative of production queries\n**Solution**: Continuous test set updates from production data\n\n**Problem**: Annotator bias in relevance judgments\n**Solution**: Multiple annotators, clear guidelines, bias training\n\n### 2. Metric Gaming\n\n**Problem**: Optimizing for metrics rather than user satisfaction\n**Solution**: Multiple complementary metrics, regular metric validation\n\n**Problem**: Overfitting to evaluation set\n**Solution**: Hold-out validation sets, temporal splits\n\n### 3. Scale Challenges\n\n**Problem**: Evaluation becomes too expensive at scale\n**Solution**: Sampling strategies, automated metrics, efficient tooling\n\n**Problem**: Human evaluation bottlenecks\n**Solution**: Active learning for annotation, LLM-as-judge validation\n\n## Future Directions\n\n### 1. Advanced Metrics\n\n- **Semantic Coherence**: Measuring logical flow in generated answers\n- **Factual Consistency**: Cross-document fact verification\n- **Personalization Quality**: User-specific relevance assessment\n- **Multimodal Evaluation**: Text, image, audio integration metrics\n\n### 2. Automated Evaluation\n\n- **LLM-as-Judge**: Using large language models for quality assessment\n- **Adversarial Testing**: Systematic stress testing of RAG systems\n- **Causal Evaluation**: Understanding why systems fail\n- **Real-time Adaptation**: Dynamic metric adjustment based on context\n\n### 3. Holistic Assessment\n\n- **User Journey Evaluation**: Multi-turn conversation quality\n- **Task Success Measurement**: Goal completion rather than single query\n- **Temporal Consistency**: Performance stability over time\n- **Fairness and Bias**: Systematic bias detection and measurement\n\n## Conclusion\n\nEffective RAG evaluation requires a multi-faceted approach combining automated metrics, human judgment, and continuous monitoring. The key principles are:\n\n1. **Comprehensive Coverage**: Evaluate all pipeline components\n2. **Multiple Perspectives**: Combine different evaluation methodologies \n3. **Continuous Improvement**: Regular evaluation and iteration\n4. **Business Alignment**: Metrics should reflect actual user value\n5. **Statistical Rigor**: Proper experimental design and analysis\n\nThis framework provides the foundation for building robust, high-quality RAG systems that deliver real value to users while maintaining reliability and trustworthiness.","content_type":"text/markdown; charset=utf-8","language":"markdown","size":15137,"content_sha256":"6c625f57489db2bf9f0836dba33eb8e253c59b01bf4516f69b536bdefbb5a2f6"},{"filename":"retrieval_evaluator.py","content":"#!/usr/bin/env python3\n\"\"\"\nRetrieval Evaluator - Evaluates retrieval quality using standard IR metrics.\n\nThis script evaluates retrieval system performance using standard information retrieval\nmetrics including precision@k, recall@k, MRR, and NDCG. It uses a built-in TF-IDF\nimplementation as a baseline retrieval system.\n\nMetrics calculated:\n- Precision@K: Fraction of retrieved documents that are relevant\n- Recall@K: Fraction of relevant documents that are retrieved \n- Mean Reciprocal Rank (MRR): Average reciprocal rank of first relevant result\n- Normalized Discounted Cumulative Gain (NDCG): Ranking quality with position discount\n\nNo external dependencies - uses only Python standard library.\n\"\"\"\n\nimport argparse\nimport json\nimport math\nimport os\nimport re\nfrom collections import Counter, defaultdict\nfrom pathlib import Path\nfrom typing import Dict, List, Tuple, Set, Any, Optional\n\n\nclass Document:\n \"\"\"Represents a document in the corpus.\"\"\"\n \n def __init__(self, doc_id: str, title: str, content: str, path: str = \"\"):\n self.doc_id = doc_id\n self.title = title\n self.content = content\n self.path = path\n self.tokens = self._tokenize(content)\n self.token_count = len(self.tokens)\n \n def _tokenize(self, text: str) -> List[str]:\n \"\"\"Simple tokenization - split on whitespace and punctuation.\"\"\"\n # Convert to lowercase and extract words\n tokens = re.findall(r'\\b[a-zA-Z0-9]+\\b', text.lower())\n return tokens\n \n def __str__(self):\n return f\"Document({self.doc_id}, '{self.title[:50]}...', {self.token_count} tokens)\"\n\n\nclass TFIDFRetriever:\n \"\"\"TF-IDF based retrieval system - no external dependencies.\"\"\"\n \n def __init__(self, documents: List[Document]):\n self.documents = {doc.doc_id: doc for doc in documents}\n self.doc_ids = list(self.documents.keys())\n self.vocabulary = set()\n self.tf_scores = {} # doc_id -> {term: tf_score}\n self.df_scores = {} # term -> document_frequency\n self.idf_scores = {} # term -> idf_score\n self._build_index()\n \n def _build_index(self):\n \"\"\"Build TF-IDF index from documents.\"\"\"\n print(f\"Building TF-IDF index for {len(self.documents)} documents...\")\n \n # Calculate term frequencies and build vocabulary\n for doc_id, doc in self.documents.items():\n term_counts = Counter(doc.tokens)\n doc_length = len(doc.tokens)\n \n # Calculate TF scores (term_count / doc_length)\n tf_scores = {}\n for term, count in term_counts.items():\n tf_scores[term] = count / doc_length if doc_length > 0 else 0\n self.vocabulary.add(term)\n \n self.tf_scores[doc_id] = tf_scores\n \n # Calculate document frequencies\n for term in self.vocabulary:\n df = sum(1 for doc in self.documents.values() if term in doc.tokens)\n self.df_scores[term] = df\n \n # Calculate IDF scores: log(N / df)\n num_docs = len(self.documents)\n for term, df in self.df_scores.items():\n self.idf_scores[term] = math.log(num_docs / df) if df > 0 else 0\n \n def search(self, query: str, k: int = 10) -> List[Tuple[str, float]]:\n \"\"\"Search for documents matching the query using TF-IDF similarity.\"\"\"\n query_tokens = re.findall(r'\\b[a-zA-Z0-9]+\\b', query.lower())\n if not query_tokens:\n return []\n \n # Calculate query TF scores\n query_tf = Counter(query_tokens)\n query_length = len(query_tokens)\n \n # Calculate TF-IDF similarity for each document\n scores = {}\n for doc_id in self.doc_ids:\n score = self._calculate_similarity(query_tf, query_length, doc_id)\n if score > 0:\n scores[doc_id] = score\n \n # Sort by score and return top k\n sorted_results = sorted(scores.items(), key=lambda x: x[1], reverse=True)\n return sorted_results[:k]\n \n def _calculate_similarity(self, query_tf: Counter, query_length: int, doc_id: str) -> float:\n \"\"\"Calculate cosine similarity between query and document using TF-IDF.\"\"\"\n doc_tf = self.tf_scores[doc_id]\n \n # Calculate TF-IDF vectors\n query_vector = []\n doc_vector = []\n \n # Only consider terms that appear in both query and document\n common_terms = set(query_tf.keys()) & set(doc_tf.keys())\n \n if not common_terms:\n return 0.0\n \n for term in common_terms:\n # Query TF-IDF\n q_tf = query_tf[term] / query_length\n q_tfidf = q_tf * self.idf_scores.get(term, 0)\n query_vector.append(q_tfidf)\n \n # Document TF-IDF\n d_tfidf = doc_tf[term] * self.idf_scores.get(term, 0)\n doc_vector.append(d_tfidf)\n \n # Cosine similarity\n dot_product = sum(q * d for q, d in zip(query_vector, doc_vector))\n query_norm = math.sqrt(sum(q * q for q in query_vector))\n doc_norm = math.sqrt(sum(d * d for d in doc_vector))\n \n if query_norm == 0 or doc_norm == 0:\n return 0.0\n \n return dot_product / (query_norm * doc_norm)\n\n\nclass RetrievalEvaluator:\n \"\"\"Evaluates retrieval system performance using standard IR metrics.\"\"\"\n \n def __init__(self):\n self.metrics = {}\n \n def evaluate(self, queries: List[Dict[str, Any]], ground_truth: Dict[str, List[str]], \n retriever: TFIDFRetriever, k_values: List[int] = None) -> Dict[str, Any]:\n \"\"\"Evaluate retrieval performance.\"\"\"\n k_values = k_values or [1, 3, 5, 10]\n \n print(f\"Evaluating retrieval performance for {len(queries)} queries...\")\n \n query_results = []\n all_precision_at_k = {k: [] for k in k_values}\n all_recall_at_k = {k: [] for k in k_values}\n all_ndcg_at_k = {k: [] for k in k_values}\n reciprocal_ranks = []\n \n for query_data in queries:\n query_id = query_data['id']\n query_text = query_data['query']\n \n # Get ground truth for this query\n relevant_docs = set(ground_truth.get(query_id, []))\n \n if not relevant_docs:\n print(f\"Warning: No ground truth found for query {query_id}\")\n continue\n \n # Retrieve documents\n max_k = max(k_values)\n results = retriever.search(query_text, max_k)\n retrieved_doc_ids = [doc_id for doc_id, _ in results]\n \n # Calculate metrics for this query\n query_metrics = {}\n \n # Precision@K and Recall@K\n for k in k_values:\n retrieved_at_k = set(retrieved_doc_ids[:k])\n relevant_retrieved = retrieved_at_k & relevant_docs\n \n precision = len(relevant_retrieved) / len(retrieved_at_k) if retrieved_at_k else 0\n recall = len(relevant_retrieved) / len(relevant_docs) if relevant_docs else 0\n \n query_metrics[f'precision@{k}'] = precision\n query_metrics[f'recall@{k}'] = recall\n \n all_precision_at_k[k].append(precision)\n all_recall_at_k[k].append(recall)\n \n # Mean Reciprocal Rank (MRR)\n reciprocal_rank = self._calculate_reciprocal_rank(retrieved_doc_ids, relevant_docs)\n query_metrics['reciprocal_rank'] = reciprocal_rank\n reciprocal_ranks.append(reciprocal_rank)\n \n # NDCG@K\n for k in k_values:\n ndcg = self._calculate_ndcg(retrieved_doc_ids[:k], relevant_docs)\n query_metrics[f'ndcg@{k}'] = ndcg\n all_ndcg_at_k[k].append(ndcg)\n \n # Store query-level results\n query_results.append({\n 'query_id': query_id,\n 'query': query_text,\n 'relevant_count': len(relevant_docs),\n 'retrieved_count': len(retrieved_doc_ids),\n 'metrics': query_metrics,\n 'retrieved_docs': results[:5], # Top 5 for analysis\n 'relevant_docs': list(relevant_docs)\n })\n \n # Calculate aggregate metrics\n aggregate_metrics = {}\n \n for k in k_values:\n aggregate_metrics[f'mean_precision@{k}'] = self._safe_mean(all_precision_at_k[k])\n aggregate_metrics[f'mean_recall@{k}'] = self._safe_mean(all_recall_at_k[k])\n aggregate_metrics[f'mean_ndcg@{k}'] = self._safe_mean(all_ndcg_at_k[k])\n \n aggregate_metrics['mean_reciprocal_rank'] = self._safe_mean(reciprocal_ranks)\n \n # Failure analysis\n failure_analysis = self._analyze_failures(query_results)\n \n return {\n 'aggregate_metrics': aggregate_metrics,\n 'query_results': query_results,\n 'failure_analysis': failure_analysis,\n 'evaluation_summary': self._generate_summary(aggregate_metrics, len(queries))\n }\n \n def _calculate_reciprocal_rank(self, retrieved_docs: List[str], relevant_docs: Set[str]) -> float:\n \"\"\"Calculate reciprocal rank - 1/rank of first relevant document.\"\"\"\n for i, doc_id in enumerate(retrieved_docs):\n if doc_id in relevant_docs:\n return 1.0 / (i + 1)\n return 0.0\n \n def _calculate_ndcg(self, retrieved_docs: List[str], relevant_docs: Set[str]) -> float:\n \"\"\"Calculate Normalized Discounted Cumulative Gain.\"\"\"\n if not retrieved_docs:\n return 0.0\n \n # DCG calculation\n dcg = 0.0\n for i, doc_id in enumerate(retrieved_docs):\n relevance = 1 if doc_id in relevant_docs else 0\n dcg += relevance / math.log2(i + 2) # +2 because log2(1) = 0\n \n # IDCG calculation (ideal DCG)\n ideal_relevances = [1] * min(len(relevant_docs), len(retrieved_docs))\n idcg = sum(rel / math.log2(i + 2) for i, rel in enumerate(ideal_relevances))\n \n return dcg / idcg if idcg > 0 else 0.0\n \n def _safe_mean(self, values: List[float]) -> float:\n \"\"\"Calculate mean, handling empty lists.\"\"\"\n return sum(values) / len(values) if values else 0.0\n \n def _analyze_failures(self, query_results: List[Dict[str, Any]]) -> Dict[str, Any]:\n \"\"\"Analyze common failure patterns.\"\"\"\n total_queries = len(query_results)\n \n # Identify queries with poor performance\n poor_precision_queries = []\n poor_recall_queries = []\n zero_results_queries = []\n \n for result in query_results:\n metrics = result['metrics']\n \n if metrics.get('precision@5', 0) \u003c 0.2:\n poor_precision_queries.append(result)\n \n if metrics.get('recall@5', 0) \u003c 0.3:\n poor_recall_queries.append(result)\n \n if result['retrieved_count'] == 0:\n zero_results_queries.append(result)\n \n # Analyze query characteristics\n query_length_analysis = self._analyze_query_lengths(query_results)\n \n return {\n 'poor_precision_count': len(poor_precision_queries),\n 'poor_recall_count': len(poor_recall_queries),\n 'zero_results_count': len(zero_results_queries),\n 'poor_precision_examples': poor_precision_queries[:3],\n 'poor_recall_examples': poor_recall_queries[:3],\n 'query_length_analysis': query_length_analysis,\n 'common_failure_patterns': self._identify_failure_patterns(query_results)\n }\n \n def _analyze_query_lengths(self, query_results: List[Dict[str, Any]]) -> Dict[str, Any]:\n \"\"\"Analyze relationship between query length and performance.\"\"\"\n short_queries = [] # \u003c= 3 words\n medium_queries = [] # 4-7 words \n long_queries = [] # >= 8 words\n \n for result in query_results:\n query_length = len(result['query'].split())\n precision = result['metrics'].get('precision@5', 0)\n \n if query_length \u003c= 3:\n short_queries.append(precision)\n elif query_length \u003c= 7:\n medium_queries.append(precision)\n else:\n long_queries.append(precision)\n \n return {\n 'short_queries': {\n 'count': len(short_queries),\n 'avg_precision@5': self._safe_mean(short_queries)\n },\n 'medium_queries': {\n 'count': len(medium_queries),\n 'avg_precision@5': self._safe_mean(medium_queries)\n },\n 'long_queries': {\n 'count': len(long_queries),\n 'avg_precision@5': self._safe_mean(long_queries)\n }\n }\n \n def _identify_failure_patterns(self, query_results: List[Dict[str, Any]]) -> List[str]:\n \"\"\"Identify common patterns in failed queries.\"\"\"\n patterns = []\n \n # Check for vocabulary mismatch\n vocab_mismatch_count = 0\n for result in query_results:\n if result['metrics'].get('precision@1', 0) == 0 and result['retrieved_count'] > 0:\n vocab_mismatch_count += 1\n \n if vocab_mismatch_count > len(query_results) * 0.2:\n patterns.append(f\"Vocabulary mismatch: {vocab_mismatch_count} queries may have vocabulary mismatch issues\")\n \n # Check for specificity issues\n zero_results = sum(1 for r in query_results if r['retrieved_count'] == 0)\n if zero_results > len(query_results) * 0.1:\n patterns.append(f\"Query specificity: {zero_results} queries returned no results (may be too specific)\")\n \n # Check for recall issues\n low_recall = sum(1 for r in query_results if r['metrics'].get('recall@10', 0) \u003c 0.5)\n if low_recall > len(query_results) * 0.3:\n patterns.append(f\"Low recall: {low_recall} queries have recall@10 \u003c 0.5 (missing relevant documents)\")\n \n return patterns\n \n def _generate_summary(self, metrics: Dict[str, float], num_queries: int) -> str:\n \"\"\"Generate human-readable evaluation summary.\"\"\"\n summary = f\"Evaluation Summary ({num_queries} queries):\\n\"\n summary += f\"{'='*50}\\n\"\n \n # Key metrics\n p1 = metrics.get('mean_precision@1', 0)\n p5 = metrics.get('mean_precision@5', 0)\n r5 = metrics.get('mean_recall@5', 0)\n mrr = metrics.get('mean_reciprocal_rank', 0)\n ndcg5 = metrics.get('mean_ndcg@5', 0)\n \n summary += f\"Precision@1: {p1:.3f} ({p1*100:.1f}%)\\n\"\n summary += f\"Precision@5: {p5:.3f} ({p5*100:.1f}%)\\n\"\n summary += f\"Recall@5: {r5:.3f} ({r5*100:.1f}%)\\n\"\n summary += f\"MRR: {mrr:.3f}\\n\"\n summary += f\"NDCG@5: {ndcg5:.3f}\\n\"\n \n # Performance assessment\n summary += f\"\\nPerformance Assessment:\\n\"\n if p1 >= 0.7:\n summary += \"✓ Excellent precision - most queries return relevant results first\\n\"\n elif p1 >= 0.5:\n summary += \"○ Good precision - many queries return relevant results first\\n\"\n else:\n summary += \"✗ Poor precision - few queries return relevant results first\\n\"\n \n if r5 >= 0.8:\n summary += \"✓ Excellent recall - finding most relevant documents\\n\"\n elif r5 >= 0.6:\n summary += \"○ Good recall - finding many relevant documents\\n\"\n else:\n summary += \"✗ Poor recall - missing many relevant documents\\n\"\n \n return summary\n\n\ndef load_queries(file_path: str) -> List[Dict[str, Any]]:\n \"\"\"Load queries from JSON file.\"\"\"\n with open(file_path, 'r', encoding='utf-8') as f:\n data = json.load(f)\n \n # Handle different JSON formats\n if isinstance(data, list):\n return data\n elif 'queries' in data:\n return data['queries']\n else:\n raise ValueError(\"Invalid query file format. Expected list of queries or {'queries': [...]}.\")\n\n\ndef load_ground_truth(file_path: str) -> Dict[str, List[str]]:\n \"\"\"Load ground truth relevance judgments.\"\"\"\n with open(file_path, 'r', encoding='utf-8') as f:\n data = json.load(f)\n \n # Handle different JSON formats\n if isinstance(data, dict):\n # Convert all values to lists if they aren't already\n return {k: v if isinstance(v, list) else [v] for k, v in data.items()}\n else:\n raise ValueError(\"Invalid ground truth format. Expected dict mapping query_id -> relevant_doc_ids.\")\n\n\ndef load_corpus(directory: str, extensions: List[str] = None) -> List[Document]:\n \"\"\"Load document corpus from directory.\"\"\"\n extensions = extensions or ['.txt', '.md', '.markdown']\n documents = []\n \n corpus_path = Path(directory)\n if not corpus_path.exists():\n raise FileNotFoundError(f\"Corpus directory not found: {directory}\")\n \n for file_path in corpus_path.rglob('*'):\n if file_path.is_file() and file_path.suffix.lower() in extensions:\n try:\n with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:\n content = f.read()\n \n if content.strip():\n # Use filename (without extension) as doc_id\n doc_id = file_path.stem\n title = file_path.name\n \n doc = Document(doc_id, title, content, str(file_path))\n documents.append(doc)\n \n except Exception as e:\n print(f\"Warning: Could not read {file_path}: {e}\")\n \n if not documents:\n raise ValueError(f\"No valid documents found in {directory}\")\n \n print(f\"Loaded {len(documents)} documents from corpus\")\n return documents\n\n\ndef generate_recommendations(evaluation_results: Dict[str, Any]) -> List[str]:\n \"\"\"Generate improvement recommendations based on evaluation results.\"\"\"\n recommendations = []\n \n metrics = evaluation_results['aggregate_metrics']\n failure_analysis = evaluation_results['failure_analysis']\n \n # Precision-based recommendations\n p1 = metrics.get('mean_precision@1', 0)\n p5 = metrics.get('mean_precision@5', 0)\n \n if p1 \u003c 0.3:\n recommendations.append(\"LOW PRECISION: Consider implementing query expansion or reranking to improve result quality.\")\n \n if p5 \u003c 0.4:\n recommendations.append(\"RANKING ISSUES: Current ranking may not prioritize relevant documents. Consider BM25 or learning-to-rank models.\")\n \n # Recall-based recommendations\n r5 = metrics.get('mean_recall@5', 0)\n r10 = metrics.get('mean_recall@10', 0)\n \n if r5 \u003c 0.5:\n recommendations.append(\"LOW RECALL: Consider query expansion techniques (synonyms, related terms) to find more relevant documents.\")\n \n if r10 - r5 > 0.2:\n recommendations.append(\"RANKING DEPTH: Many relevant documents found in positions 6-10. Consider increasing default result count.\")\n \n # MRR-based recommendations\n mrr = metrics.get('mean_reciprocal_rank', 0)\n if mrr \u003c 0.4:\n recommendations.append(\"POOR RANKING: First relevant result appears late in rankings. Implement result reranking.\")\n \n # Failure pattern recommendations\n zero_results = failure_analysis.get('zero_results_count', 0)\n total_queries = len(evaluation_results['query_results'])\n \n if zero_results > total_queries * 0.1:\n recommendations.append(\"COVERAGE ISSUES: Many queries return no results. Check for vocabulary mismatch or missing content.\")\n \n # Query length analysis\n query_analysis = failure_analysis.get('query_length_analysis', {})\n short_perf = query_analysis.get('short_queries', {}).get('avg_precision@5', 0)\n long_perf = query_analysis.get('long_queries', {}).get('avg_precision@5', 0)\n \n if short_perf \u003c 0.3:\n recommendations.append(\"SHORT QUERY ISSUES: Brief queries perform poorly. Consider query completion or suggestion features.\")\n \n if long_perf > short_perf + 0.2:\n recommendations.append(\"QUERY PROCESSING: Longer queries perform better. Consider query parsing to extract key terms.\")\n \n # General recommendations\n if not recommendations:\n recommendations.append(\"GOOD PERFORMANCE: System performs well overall. Consider A/B testing incremental improvements.\")\n \n return recommendations\n\n\ndef main():\n \"\"\"Main function with command-line interface.\"\"\"\n parser = argparse.ArgumentParser(description='Evaluate retrieval system performance')\n parser.add_argument('queries', help='JSON file containing queries')\n parser.add_argument('corpus', help='Directory containing document corpus')\n parser.add_argument('ground_truth', help='JSON file containing ground truth relevance judgments')\n parser.add_argument('--output', '-o', help='Output file for results (JSON format)')\n parser.add_argument('--k-values', nargs='+', type=int, default=[1, 3, 5, 10], \n help='K values for precision@k, recall@k, NDCG@k evaluation')\n parser.add_argument('--extensions', nargs='+', default=['.txt', '.md', '.markdown'], \n help='File extensions to include from corpus')\n parser.add_argument('--verbose', '-v', action='store_true', help='Verbose output')\n \n args = parser.parse_args()\n \n try:\n # Load data\n print(\"Loading evaluation data...\")\n queries = load_queries(args.queries)\n ground_truth = load_ground_truth(args.ground_truth)\n documents = load_corpus(args.corpus, args.extensions)\n \n print(f\"Loaded {len(queries)} queries, {len(documents)} documents, ground truth for {len(ground_truth)} queries\")\n \n # Build retrieval system\n retriever = TFIDFRetriever(documents)\n \n # Run evaluation\n evaluator = RetrievalEvaluator()\n results = evaluator.evaluate(queries, ground_truth, retriever, args.k_values)\n \n # Generate recommendations\n recommendations = generate_recommendations(results)\n results['recommendations'] = recommendations\n \n # Save results\n if args.output:\n with open(args.output, 'w') as f:\n json.dump(results, f, indent=2)\n print(f\"Results saved to {args.output}\")\n \n # Print summary\n print(\"\\n\" + results['evaluation_summary'])\n \n print(\"\\nRecommendations:\")\n for i, rec in enumerate(recommendations, 1):\n print(f\"{i}. {rec}\")\n \n if args.verbose:\n print(f\"\\nDetailed Metrics:\")\n for metric, value in results['aggregate_metrics'].items():\n print(f\" {metric}: {value:.4f}\")\n \n print(f\"\\nFailure Analysis:\")\n fa = results['failure_analysis']\n print(f\" Poor precision queries: {fa['poor_precision_count']}\")\n print(f\" Poor recall queries: {fa['poor_recall_count']}\")\n print(f\" Zero result queries: {fa['zero_results_count']}\")\n \n except Exception as e:\n print(f\"Error: {e}\")\n return 1\n \n return 0\n\n\nif __name__ == '__main__':\n exit(main())","content_type":"text/x-python; charset=utf-8","language":"python","size":23603,"content_sha256":"43b05ec3861a929c0317bac8e9605108d820c16683a10c90acc44f4243cf873d"}],"content_json":{"type":"doc","content":[{"type":"heading","attrs":{"level":1},"content":[{"text":"RAG Architect - POWERFUL","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Overview","type":"text"}]},{"type":"paragraph","content":[{"text":"The RAG (Retrieval-Augmented Generation) Architect skill provides comprehensive tools and knowledge for designing, implementing, and optimizing production-grade RAG pipelines. This skill covers the entire RAG ecosystem from document chunking strategies to evaluation frameworks, enabling you to build scalable, efficient, and accurate retrieval systems.","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Core Competencies","type":"text"}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"1. Document Processing & Chunking Strategies","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Fixed-Size Chunking","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Character-based chunking","type":"text","marks":[{"type":"strong"}]},{"text":": Simple splitting by character count (e.g., 512, 1024, 2048 chars)","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Token-based chunking","type":"text","marks":[{"type":"strong"}]},{"text":": Splitting by token count to respect model limits","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Overlap strategies","type":"text","marks":[{"type":"strong"}]},{"text":": 10-20% overlap to maintain context continuity","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pros","type":"text","marks":[{"type":"strong"}]},{"text":": Predictable chunk sizes, simple implementation, consistent processing time","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": May break semantic units, context boundaries ignored","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Uniform documents, when consistent chunk sizes are critical","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Sentence-Based Chunking","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Sentence boundary detection","type":"text","marks":[{"type":"strong"}]},{"text":": Using NLTK, spaCy, or regex patterns","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Sentence grouping","type":"text","marks":[{"type":"strong"}]},{"text":": Combining sentences until size threshold is reached","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Paragraph preservation","type":"text","marks":[{"type":"strong"}]},{"text":": Avoiding mid-paragraph splits when possible","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pros","type":"text","marks":[{"type":"strong"}]},{"text":": Preserves natural language boundaries, better readability","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Variable chunk sizes, potential for very short/long chunks","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Narrative text, articles, books","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Paragraph-Based Chunking","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Paragraph detection","type":"text","marks":[{"type":"strong"}]},{"text":": Double newlines, HTML tags, markdown formatting","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Hierarchical splitting","type":"text","marks":[{"type":"strong"}]},{"text":": Respecting document structure (sections, subsections)","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Size balancing","type":"text","marks":[{"type":"strong"}]},{"text":": Merging small paragraphs, splitting large ones","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pros","type":"text","marks":[{"type":"strong"}]},{"text":": Preserves logical document structure, maintains topic coherence","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Highly variable sizes, may create very large chunks","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Structured documents, technical documentation","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Semantic Chunking","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Topic modeling","type":"text","marks":[{"type":"strong"}]},{"text":": Using TF-IDF, embeddings similarity for topic detection","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Heading-aware splitting","type":"text","marks":[{"type":"strong"}]},{"text":": Respecting document hierarchy (H1, H2, H3)","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Content-based boundaries","type":"text","marks":[{"type":"strong"}]},{"text":": Detecting topic shifts using semantic similarity","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pros","type":"text","marks":[{"type":"strong"}]},{"text":": Maintains semantic coherence, respects document structure","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Complex implementation, computationally expensive","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Long-form content, technical manuals, research papers","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Recursive Chunking","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Hierarchical approach","type":"text","marks":[{"type":"strong"}]},{"text":": Try larger chunks first, recursively split if needed","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Multi-level splitting","type":"text","marks":[{"type":"strong"}]},{"text":": Different strategies at different levels","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Size optimization","type":"text","marks":[{"type":"strong"}]},{"text":": Minimize number of chunks while respecting size limits","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pros","type":"text","marks":[{"type":"strong"}]},{"text":": Optimal chunk utilization, preserves context when possible","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Complex logic, potential performance overhead","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Mixed content types, when chunk count optimization is important","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Document-Aware Chunking","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"File type detection","type":"text","marks":[{"type":"strong"}]},{"text":": PDF pages, Word sections, HTML elements","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Metadata preservation","type":"text","marks":[{"type":"strong"}]},{"text":": Headers, footers, page numbers, sections","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Table and image handling","type":"text","marks":[{"type":"strong"}]},{"text":": Special processing for non-text elements","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pros","type":"text","marks":[{"type":"strong"}]},{"text":": Preserves document structure and metadata","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Format-specific implementation required","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Multi-format document collections, when metadata is important","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"2. Embedding Model Selection","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Dimension Considerations","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"128-256 dimensions","type":"text","marks":[{"type":"strong"}]},{"text":": Fast retrieval, lower memory usage, suitable for simple domains","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"512-768 dimensions","type":"text","marks":[{"type":"strong"}]},{"text":": Balanced performance, good for most applications","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"1024-1536 dimensions","type":"text","marks":[{"type":"strong"}]},{"text":": High quality, better for complex domains, higher cost","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"2048+ dimensions","type":"text","marks":[{"type":"strong"}]},{"text":": Maximum quality, specialized use cases, significant resources","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Speed vs Quality Tradeoffs","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Fast models","type":"text","marks":[{"type":"strong"}]},{"text":": sentence-transformers/all-MiniLM-L6-v2 (384 dim, ~14k tokens/sec)","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Balanced models","type":"text","marks":[{"type":"strong"}]},{"text":": sentence-transformers/all-mpnet-base-v2 (768 dim, ~2.8k tokens/sec)","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Quality models","type":"text","marks":[{"type":"strong"}]},{"text":": text-embedding-ada-002 (1536 dim, OpenAI API)","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Specialized models","type":"text","marks":[{"type":"strong"}]},{"text":": Domain-specific fine-tuned models","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Model Categories","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"General purpose","type":"text","marks":[{"type":"strong"}]},{"text":": all-MiniLM, all-mpnet, Universal Sentence Encoder","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Code embeddings","type":"text","marks":[{"type":"strong"}]},{"text":": CodeBERT, GraphCodeBERT, CodeT5","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Scientific text","type":"text","marks":[{"type":"strong"}]},{"text":": SciBERT, BioBERT, ClinicalBERT","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Multilingual","type":"text","marks":[{"type":"strong"}]},{"text":": LaBSE, multilingual-e5, paraphrase-multilingual","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"3. Vector Database Selection","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Pinecone","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Managed service","type":"text","marks":[{"type":"strong"}]},{"text":": Fully hosted, auto-scaling","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Features","type":"text","marks":[{"type":"strong"}]},{"text":": Metadata filtering, hybrid search, real-time updates","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pricing","type":"text","marks":[{"type":"strong"}]},{"text":": $70/month for 1M vectors (1536 dim), pay-per-use scaling","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Production applications, when managed service is preferred","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Vendor lock-in, costs can scale quickly","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Weaviate","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Open source","type":"text","marks":[{"type":"strong"}]},{"text":": Self-hosted or cloud options available","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Features","type":"text","marks":[{"type":"strong"}]},{"text":": GraphQL API, multi-modal search, automatic vectorization","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Scaling","type":"text","marks":[{"type":"strong"}]},{"text":": Horizontal scaling, HNSW indexing","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Complex data types, when GraphQL API is preferred","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Learning curve, requires infrastructure management","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Qdrant","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Rust-based","type":"text","marks":[{"type":"strong"}]},{"text":": High performance, low memory footprint","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Features","type":"text","marks":[{"type":"strong"}]},{"text":": Payload filtering, clustering, distributed deployment","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"API","type":"text","marks":[{"type":"strong"}]},{"text":": REST and gRPC interfaces","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": High-performance requirements, resource-constrained environments","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Smaller community, fewer integrations","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Chroma","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Embedded database","type":"text","marks":[{"type":"strong"}]},{"text":": SQLite-based, easy local development","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Features","type":"text","marks":[{"type":"strong"}]},{"text":": Collections, metadata filtering, persistence","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Scaling","type":"text","marks":[{"type":"strong"}]},{"text":": Limited, suitable for prototyping and small deployments","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": Development, testing, small-scale applications","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Not suitable for production scale","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"pgvector (PostgreSQL)","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"SQL integration","type":"text","marks":[{"type":"strong"}]},{"text":": Leverage existing PostgreSQL infrastructure","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Features","type":"text","marks":[{"type":"strong"}]},{"text":": ACID compliance, joins with relational data, mature ecosystem","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Performance","type":"text","marks":[{"type":"strong"}]},{"text":": ivfflat and HNSW indexing, parallel query processing","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Best for","type":"text","marks":[{"type":"strong"}]},{"text":": When you already use PostgreSQL, need ACID compliance","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cons","type":"text","marks":[{"type":"strong"}]},{"text":": Requires PostgreSQL expertise, less specialized than purpose-built DBs","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"4. Retrieval Strategies","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Dense Retrieval","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Semantic similarity","type":"text","marks":[{"type":"strong"}]},{"text":": Using embedding cosine similarity","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Advantages","type":"text","marks":[{"type":"strong"}]},{"text":": Captures semantic meaning, handles paraphrasing well","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Limitations","type":"text","marks":[{"type":"strong"}]},{"text":": May miss exact keyword matches, requires good embeddings","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Implementation","type":"text","marks":[{"type":"strong"}]},{"text":": Vector similarity search with k-NN or ANN algorithms","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Sparse Retrieval","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Keyword-based","type":"text","marks":[{"type":"strong"}]},{"text":": TF-IDF, BM25, Elasticsearch","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Advantages","type":"text","marks":[{"type":"strong"}]},{"text":": Exact keyword matching, interpretable results","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Limitations","type":"text","marks":[{"type":"strong"}]},{"text":": Misses semantic similarity, vulnerable to vocabulary mismatch","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Implementation","type":"text","marks":[{"type":"strong"}]},{"text":": Inverted indexes, term frequency analysis","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Hybrid Retrieval","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Combination approach","type":"text","marks":[{"type":"strong"}]},{"text":": Dense + sparse retrieval with score fusion","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Fusion strategies","type":"text","marks":[{"type":"strong"}]},{"text":": Reciprocal Rank Fusion (RRF), weighted combination","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Benefits","type":"text","marks":[{"type":"strong"}]},{"text":": Combines semantic understanding with exact matching","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Complexity","type":"text","marks":[{"type":"strong"}]},{"text":": Requires tuning fusion weights, more complex infrastructure","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Reranking","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Two-stage approach","type":"text","marks":[{"type":"strong"}]},{"text":": Initial retrieval followed by reranking","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Reranking models","type":"text","marks":[{"type":"strong"}]},{"text":": Cross-encoders, specialized reranking transformers","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Benefits","type":"text","marks":[{"type":"strong"}]},{"text":": Higher precision, can use more sophisticated models for final ranking","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Tradeoff","type":"text","marks":[{"type":"strong"}]},{"text":": Additional latency, computational cost","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"5. Query Transformation Techniques","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"HyDE (Hypothetical Document Embeddings)","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Approach","type":"text","marks":[{"type":"strong"}]},{"text":": Generate hypothetical answer, embed answer instead of query","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Benefits","type":"text","marks":[{"type":"strong"}]},{"text":": Improves retrieval by matching document style rather than query style","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Implementation","type":"text","marks":[{"type":"strong"}]},{"text":": Use LLM to generate hypothetical document, embed that","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Use cases","type":"text","marks":[{"type":"strong"}]},{"text":": When queries and documents have different styles","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Multi-Query Generation","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Approach","type":"text","marks":[{"type":"strong"}]},{"text":": Generate multiple query variations, retrieve for each, merge results","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Benefits","type":"text","marks":[{"type":"strong"}]},{"text":": Increases recall, handles query ambiguity","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Implementation","type":"text","marks":[{"type":"strong"}]},{"text":": LLM generates 3-5 query variations, deduplicate results","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Considerations","type":"text","marks":[{"type":"strong"}]},{"text":": Higher cost and latency due to multiple retrievals","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Step-Back Prompting","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Approach","type":"text","marks":[{"type":"strong"}]},{"text":": Generate broader, more general version of specific query","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Benefits","type":"text","marks":[{"type":"strong"}]},{"text":": Retrieves more general context that helps answer specific questions","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Implementation","type":"text","marks":[{"type":"strong"}]},{"text":": Transform \"What is the capital of France?\" to \"What are European capitals?\"","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Use cases","type":"text","marks":[{"type":"strong"}]},{"text":": When specific questions need general context","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"6. Context Window Optimization","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Dynamic Context Assembly","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Relevance-based ordering","type":"text","marks":[{"type":"strong"}]},{"text":": Most relevant chunks first","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Diversity optimization","type":"text","marks":[{"type":"strong"}]},{"text":": Avoid redundant information","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Token budget management","type":"text","marks":[{"type":"strong"}]},{"text":": Fit within model context limits","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Hierarchical inclusion","type":"text","marks":[{"type":"strong"}]},{"text":": Include summaries before detailed chunks","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Context Compression","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Summarization","type":"text","marks":[{"type":"strong"}]},{"text":": Compress less relevant chunks while preserving key information","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Key information extraction","type":"text","marks":[{"type":"strong"}]},{"text":": Extract only relevant facts/entities","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Template-based compression","type":"text","marks":[{"type":"strong"}]},{"text":": Use structured formats to reduce token usage","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Selective inclusion","type":"text","marks":[{"type":"strong"}]},{"text":": Include only chunks above relevance threshold","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"7. Evaluation Frameworks","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Faithfulness Metrics","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Definition","type":"text","marks":[{"type":"strong"}]},{"text":": How well generated answers are grounded in retrieved context","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Measurement","type":"text","marks":[{"type":"strong"}]},{"text":": Fact verification against source documents","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Implementation","type":"text","marks":[{"type":"strong"}]},{"text":": NLI models to check entailment between answer and context","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Threshold","type":"text","marks":[{"type":"strong"}]},{"text":": >90% for production systems","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Relevance Metrics","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Context relevance","type":"text","marks":[{"type":"strong"}]},{"text":": How relevant retrieved chunks are to the query","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Answer relevance","type":"text","marks":[{"type":"strong"}]},{"text":": How well the answer addresses the original question","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Measurement","type":"text","marks":[{"type":"strong"}]},{"text":": Embedding similarity, human evaluation, LLM-as-judge","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Targets","type":"text","marks":[{"type":"strong"}]},{"text":": Context relevance >0.8, Answer relevance >0.85","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Context Precision & Recall","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Precision@K","type":"text","marks":[{"type":"strong"}]},{"text":": Percentage of top-K results that are relevant","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Recall@K","type":"text","marks":[{"type":"strong"}]},{"text":": Percentage of relevant documents found in top-K results","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Mean Reciprocal Rank (MRR)","type":"text","marks":[{"type":"strong"}]},{"text":": Average of reciprocal ranks of first relevant result","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"NDCG@K","type":"text","marks":[{"type":"strong"}]},{"text":": Normalized Discounted Cumulative Gain at K","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"End-to-End Metrics","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"RAGAS","type":"text","marks":[{"type":"strong"}]},{"text":": Comprehensive RAG evaluation framework","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Correctness","type":"text","marks":[{"type":"strong"}]},{"text":": Factual accuracy of generated answers","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Completeness","type":"text","marks":[{"type":"strong"}]},{"text":": Coverage of all relevant aspects","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Consistency","type":"text","marks":[{"type":"strong"}]},{"text":": Consistency across multiple runs with same query","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"8. Production Patterns","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Caching Strategies","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Query-level caching","type":"text","marks":[{"type":"strong"}]},{"text":": Cache results for identical queries","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Semantic caching","type":"text","marks":[{"type":"strong"}]},{"text":": Cache for semantically similar queries","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Chunk-level caching","type":"text","marks":[{"type":"strong"}]},{"text":": Cache embedding computations","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Multi-level caching","type":"text","marks":[{"type":"strong"}]},{"text":": Redis for hot queries, disk for warm queries","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Streaming Retrieval","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Progressive loading","type":"text","marks":[{"type":"strong"}]},{"text":": Stream results as they become available","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Incremental generation","type":"text","marks":[{"type":"strong"}]},{"text":": Generate answers while still retrieving","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Real-time updates","type":"text","marks":[{"type":"strong"}]},{"text":": Handle document updates without full reprocessing","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Connection management","type":"text","marks":[{"type":"strong"}]},{"text":": Handle client disconnections gracefully","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Fallback Mechanisms","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Graceful degradation","type":"text","marks":[{"type":"strong"}]},{"text":": Fallback to simpler retrieval if primary fails","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cache fallbacks","type":"text","marks":[{"type":"strong"}]},{"text":": Serve stale results when retrieval is unavailable","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Alternative sources","type":"text","marks":[{"type":"strong"}]},{"text":": Multiple vector databases for redundancy","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Error handling","type":"text","marks":[{"type":"strong"}]},{"text":": Comprehensive error recovery and user communication","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"9. Cost Optimization","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Embedding Cost Management","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Batch processing","type":"text","marks":[{"type":"strong"}]},{"text":": Batch documents for embedding to reduce API costs","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Caching strategies","type":"text","marks":[{"type":"strong"}]},{"text":": Cache embeddings to avoid recomputation","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Model selection","type":"text","marks":[{"type":"strong"}]},{"text":": Balance cost vs quality for embedding models","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Update optimization","type":"text","marks":[{"type":"strong"}]},{"text":": Only re-embed changed documents","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Vector Database Optimization","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Index optimization","type":"text","marks":[{"type":"strong"}]},{"text":": Choose appropriate index types for use case","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Compression","type":"text","marks":[{"type":"strong"}]},{"text":": Use quantization to reduce storage costs","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Tiered storage","type":"text","marks":[{"type":"strong"}]},{"text":": Hot/warm/cold data strategies","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Resource scaling","type":"text","marks":[{"type":"strong"}]},{"text":": Auto-scaling based on query patterns","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Query Optimization","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Query routing","type":"text","marks":[{"type":"strong"}]},{"text":": Route simple queries to cheaper methods","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Result caching","type":"text","marks":[{"type":"strong"}]},{"text":": Avoid repeated expensive retrievals","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Batch querying","type":"text","marks":[{"type":"strong"}]},{"text":": Process multiple queries together when possible","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Smart filtering","type":"text","marks":[{"type":"strong"}]},{"text":": Use metadata filters to reduce search space","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"10. Guardrails & Safety","type":"text"}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Content Filtering","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Toxicity detection","type":"text","marks":[{"type":"strong"}]},{"text":": Filter harmful or inappropriate content","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"PII detection","type":"text","marks":[{"type":"strong"}]},{"text":": Identify and handle personally identifiable information","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Content validation","type":"text","marks":[{"type":"strong"}]},{"text":": Ensure retrieved content meets quality standards","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Source verification","type":"text","marks":[{"type":"strong"}]},{"text":": Validate document authenticity and reliability","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Query Safety","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Injection prevention","type":"text","marks":[{"type":"strong"}]},{"text":": Prevent malicious query injection attacks","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Rate limiting","type":"text","marks":[{"type":"strong"}]},{"text":": Prevent abuse and ensure fair usage","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Query validation","type":"text","marks":[{"type":"strong"}]},{"text":": Sanitize and validate user inputs","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Access controls","type":"text","marks":[{"type":"strong"}]},{"text":": Ensure users can only access authorized content","type":"text"}]}]}]},{"type":"heading","attrs":{"level":4},"content":[{"text":"Response Safety","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Hallucination detection","type":"text","marks":[{"type":"strong"}]},{"text":": Identify when model generates unsupported claims","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Confidence scoring","type":"text","marks":[{"type":"strong"}]},{"text":": Provide confidence levels for generated responses","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Source attribution","type":"text","marks":[{"type":"strong"}]},{"text":": Always provide sources for factual claims","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Uncertainty handling","type":"text","marks":[{"type":"strong"}]},{"text":": Gracefully handle cases where answer is uncertain","type":"text"}]}]}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Implementation Best Practices","type":"text"}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Development Workflow","type":"text"}]},{"type":"ordered_list","attrs":{"order":1,"listStyle":"number"},"content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Requirements gathering","type":"text","marks":[{"type":"strong"}]},{"text":": Understand use case, scale, and quality requirements","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Data analysis","type":"text","marks":[{"type":"strong"}]},{"text":": Analyze document corpus characteristics","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Prototype development","type":"text","marks":[{"type":"strong"}]},{"text":": Build minimal viable RAG pipeline","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Chunking optimization","type":"text","marks":[{"type":"strong"}]},{"text":": Test different chunking strategies","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Retrieval tuning","type":"text","marks":[{"type":"strong"}]},{"text":": Optimize retrieval parameters and thresholds","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Evaluation setup","type":"text","marks":[{"type":"strong"}]},{"text":": Implement comprehensive evaluation metrics","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Production deployment","type":"text","marks":[{"type":"strong"}]},{"text":": Scale-ready implementation with monitoring","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Monitoring & Observability","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Query analytics","type":"text","marks":[{"type":"strong"}]},{"text":": Track query patterns and performance","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Retrieval metrics","type":"text","marks":[{"type":"strong"}]},{"text":": Monitor precision, recall, and latency","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Generation quality","type":"text","marks":[{"type":"strong"}]},{"text":": Track faithfulness and relevance scores","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"System health","type":"text","marks":[{"type":"strong"}]},{"text":": Monitor database performance and availability","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Cost tracking","type":"text","marks":[{"type":"strong"}]},{"text":": Monitor embedding and vector database costs","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Maintenance & Updates","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Document refresh","type":"text","marks":[{"type":"strong"}]},{"text":": Handle new documents and updates","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Index maintenance","type":"text","marks":[{"type":"strong"}]},{"text":": Regular vector database optimization","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Model updates","type":"text","marks":[{"type":"strong"}]},{"text":": Evaluate and migrate to improved models","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Performance tuning","type":"text","marks":[{"type":"strong"}]},{"text":": Continuous optimization based on usage patterns","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Security updates","type":"text","marks":[{"type":"strong"}]},{"text":": Regular security assessments and updates","type":"text"}]}]}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Common Pitfalls & Solutions","type":"text"}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Poor Chunking Strategy","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Problem","type":"text","marks":[{"type":"strong"}]},{"text":": Chunks break mid-sentence or lose context","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Solution","type":"text","marks":[{"type":"strong"}]},{"text":": Use boundary-aware chunking with overlap","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Low Retrieval Precision","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Problem","type":"text","marks":[{"type":"strong"}]},{"text":": Retrieved chunks are not relevant to query","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Solution","type":"text","marks":[{"type":"strong"}]},{"text":": Improve embedding model, add reranking, tune similarity threshold","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"High Latency","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Problem","type":"text","marks":[{"type":"strong"}]},{"text":": Slow retrieval and generation","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Solution","type":"text","marks":[{"type":"strong"}]},{"text":": Optimize vector indexing, implement caching, use faster embedding models","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Inconsistent Quality","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Problem","type":"text","marks":[{"type":"strong"}]},{"text":": Variable answer quality across different queries","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Solution","type":"text","marks":[{"type":"strong"}]},{"text":": Implement comprehensive evaluation, add quality scoring, improve fallbacks","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Scalability Issues","type":"text"}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Problem","type":"text","marks":[{"type":"strong"}]},{"text":": System doesn't scale with increased load","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Solution","type":"text","marks":[{"type":"strong"}]},{"text":": Implement proper caching, database sharding, and auto-scaling","type":"text"}]}]}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Conclusion","type":"text"}]},{"type":"paragraph","content":[{"text":"Building effective RAG systems requires careful consideration of each component in the pipeline. The key to success is understanding the tradeoffs between different approaches and choosing the right combination of techniques for your specific use case. Start with simple approaches and gradually add sophistication based on evaluation results and production requirements.","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"This skill provides the foundation for making informed decisions throughout the RAG development lifecycle, from initial design to production deployment and ongoing maintenance.","type":"text"}]}]},"metadata":{"date":"2026-06-05","name":"rag-architect","author":"@skillopedia","source":{"stars":16818,"repo_name":"claude-skills","origin_url":"https://github.com/alirezarezvani/claude-skills/blob/HEAD/engineering/skills/rag-architect/SKILL.md","repo_owner":"alirezarezvani","body_sha256":"dd5515e5f49239b969156dd3c4524c7e2fa5c25d83bbc6e4f49803a8d4124aaf","cluster_key":"937372f45121a747beb189b66e840577feea63574c79adc0d2d955c9613ceec7","clean_bundle":{"format":"clean-skill-bundle-v1","source":"alirezarezvani/claude-skills/engineering/skills/rag-architect/SKILL.md","attachments":[{"id":"6744389d-6fb6-570d-bbd8-c98fb1999848","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/6744389d-6fb6-570d-bbd8-c98fb1999848/attachment.py","path":"chunking_optimizer.py","size":29512,"sha256":"390cf0b370e6cabbcdc367f809e455c36da4651d7fa2808938b2a958e416d0cc","contentType":"text/x-python; charset=utf-8"},{"id":"d90d16de-b33f-504b-9e02-24e3ce62e92a","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/d90d16de-b33f-504b-9e02-24e3ce62e92a/attachment.py","path":"rag_pipeline_designer.py","size":27852,"sha256":"c285db2e2642e4001d480b26809208f92f54e4c0a6dcf021f06d53b535a536f1","contentType":"text/x-python; charset=utf-8"},{"id":"83c40487-76e1-50ce-b3f7-c8186563c0bf","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/83c40487-76e1-50ce-b3f7-c8186563c0bf/attachment.md","path":"references/chunking_strategies_comparison.md","size":10181,"sha256":"9660ca53d86521338e5e168a4fae83bf553c51e33d8f556cabb6f1ed12701b13","contentType":"text/markdown; charset=utf-8"},{"id":"318ffc29-1944-57a9-b9bf-d528a0e51a6d","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/318ffc29-1944-57a9-b9bf-d528a0e51a6d/attachment.md","path":"references/embedding_model_benchmark.md","size":12170,"sha256":"033e3c0b94b2751b1289933e01791dc11415eabf2465a0f3286e122f51a93e34","contentType":"text/markdown; charset=utf-8"},{"id":"cc3dcc95-bdba-5868-a62e-d66959e02f02","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/cc3dcc95-bdba-5868-a62e-d66959e02f02/attachment.md","path":"references/rag_evaluation_framework.md","size":15137,"sha256":"6c625f57489db2bf9f0836dba33eb8e253c59b01bf4516f69b536bdefbb5a2f6","contentType":"text/markdown; charset=utf-8"},{"id":"43cf7020-ef06-5d58-a001-6d9d54589491","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/43cf7020-ef06-5d58-a001-6d9d54589491/attachment.py","path":"retrieval_evaluator.py","size":23603,"sha256":"43b05ec3861a929c0317bac8e9605108d820c16683a10c90acc44f4243cf873d","contentType":"text/x-python; charset=utf-8"}],"bundle_sha256":"d7adb5add4fe4d465cd43e304d1312c3ebcb9210345daa5705d207809e8573af","attachment_count":6,"text_attachments":6,"attachment_storage":"skillopedia-attachments-v1","binary_attachments":0,"excluded_attachments":[]},"cluster_size":3,"skill_md_path":"engineering/skills/rag-architect/SKILL.md","import_metadata":{"date":"2026-06-05","author":"@skillopedia","version":"v1","category":"web-development","category_label":"Web"},"exact_dupes_collapsed_into_this":2},"version":"v1","category":"web-development","import_tag":"clean-skills-v1","description":"Use when the user asks to design RAG pipelines, optimize retrieval strategies, choose embedding models, implement vector search, or build knowledge retrieval systems."}},"renderedAt":1782987331460}

Important: agents should read /llm.txt, /llms.txt, or /.well-known/skills.json to discover the public Skillopedia API.