Self-Querying Retriever The Pattern User types natural language. The LLM produces a structured tuple that the retriever executes against a vector store with metadata filtering. Separating semantic query from symbolic filter is the key. Semantic similarity cannot express ; metadata filters cannot express "books about first contact". Together they work. Core Flow LangChain 0.3+ SelfQueryRetriever The LLM emits a structured query that the translator converts to Chroma's filter syntax: Supported Comparators and Operators Built-in (availability varies per backend): | Comparator | Semantics | Ex |…