Self-query retriever

1Describe your metadata schema

The retriever needs to know which fields exist on each document and what they mean. Provide an array of attribute descriptors with a name, type, and short human description. Keep the descriptions specific — the LLM uses them to decide whether a clause in the user question maps to a real filter or a free-text search.

2Wire the retriever

Pass the LLM, the vector store, the document-contents description, and the attribute info into SelfQueryRetriever.fromLLM. Meridian routes the structured-query step through azure/model-router so cheap classifications stay cheap.

import { SelfQueryRetriever } from "meridian/retrievers";
import { ChatOpenAI } from "meridian/chat";

const metadata = [
  { name: "genre", type: "string", description: "Film genre" },
  { name: "year", type: "number", description: "Release year" },
  { name: "rating", type: "number", description: "Score 0-10" },
];

const retriever = SelfQueryRetriever.fromLLM({
  llm: new ChatOpenAI({ model: "azure/model-router" }),
  vectorStore,
  documentContents: "Brief film summary",
  attributeInfo: metadata,
  verbose: true,
});

const docs = await retriever.invoke(
  "sci-fi films after 2015 rated above 8"
);

3Invoke and inspect

Call retriever.invoke(query) with the raw user string. With verbose: true Meridian logs the parsed query plan so you can verify the LLM picked the right operators. Iterate on the attribute descriptions until the filters land where you expect.