How to Implement Self-Querying Retrieval in RAG

What this does

A self-querying retriever inspects the user's question, extracts any structured filters such as date ranges or categories, and applies them during retrieval. This allows natural-language queries such as "reports from 2024" or "docs about security published by the legal team" to automatically trigger metadata filters without manual filter construction.

Steps

1. Import self-querying components. The SelfQueryRetriever requires a schema and LLM for filter extraction.

   import os
   os.environ["OLLAMA_BASE_URL"] = "http://localhost:11434"
   
   from langchain_ollama import ChatOllama, OllamaEmbeddings
   from langchain_community.vectorstores import Chroma
   from langchain.chains.query_constructor.base import AttributeInfo, FunctionalMetadatas
   

2. Define metadata attributes. Provide the LLM with field names, types, and descriptions.

   metadata_field_info = [
       AttributeInfo(
           name="date",
           description="Publication date of the document in YYYY-MM-DD format",
           type="string",
       ),
       AttributeInfo(
           name="category",
           description="Topic category, one of: engineering, legal, finance",
           type="string",
       ),
       AttributeInfo(
           name="author",
           description="Author name",
           type="string",
       ),
   ]
   

3. Load documents with metadata. Ensure each document carries structured fields.

   from langchain_community.document_loaders import JSONLoader
   
   loader = JSONLoader("context/documents.json", jq_schema=".[]", content_key="text")
   docs = loader.load()
   for doc in docs:
       doc.metadata = {
           "date": doc.metadata.get("date", "2024-01-01"),
           "category": doc.metadata.get("category", "engineering"),
           "author": doc.metadata.get("author", "unknown"),
       }
   

4. Create the self-querying retriever. Point it at the vector store and LLM.

   from langchain.retrievers.self_query.base import SelfQueryRetriever
   
   embeddings = OllamaEmbeddings(model="llama3")
   db = Chroma.from_documents(docs, embeddings)
   llm = ChatOllama(model="llama3")
   
   retriever = SelfQueryRetriever.from_llm(
       llm,
       db,
       document_content_description="Technical document",
       metadata_field_info=metadata_field_info,
   )
   

5. Query with automatic filters. The LLM extracts filters from natural language.

   results = retriever.invoke("Show me legal documents from 2024")
   for r in results:
       print(r.page_content, r.metadata)
   

   Expected output: documents whose category is "legal" and whose date starts with "2024".

Verification

python -c "
from langchain.chains.query_constructor.base import AttributeInfo
info = AttributeInfo(name='date', description='date field', type='string')
print(info.name)
# Expected: date
"

Common failures

• LLM failing to extract filters. If the model misinterprets the query, adjust the description fields in AttributeInfo to be more explicit.
• Metadata missing in documents. Every document must carry the defined fields; missing keys cause retrieval errors.
• Filter values not matching schema. Date formats, category spellings, and author names must exactly match the defined metadata; inconsistent casing breaks filtering.
• Self-query returning no results. Test the filter schema independently before wiring it into the retriever.
• Version mismatch - The installed package or runtime differs from the command shown; check the version first and rerun the smallest verification command.
• Local environment drift - Another service, virtual environment, model, or path is being used; print the active binary path and configuration before changing the guide steps.

Related guides

• add-reranking-rag-pipeline
• build-basic-rag-pipeline-langchain