Coursify

Retrieval-Augmented Generation (RAG) — From Fundamentals to Production-Ready Agentic RAG Systems

Self-Query Retriever

Implement self-query retriever for natural language to structured queries. Learn how to automatically extract query filters and metadata conditions from user questions.

Learning Goals

  • Implement self-query retriever for structured queries
  • Define query schemas and attribute mappings

Self-Query Retriever

Most RAG queries are purely semantic (e.g., "Tell me about climate change"). However, real-world questions often contain structured metadata constraints (e.g., "What were the sales in 2023 for the Western region?"). Usually, you have to write manual code to parse these filters. Self-Query Retriever automates this by using an LLM to "query itself"—it translates the user's natural language into a structured query that filters metadata and performs a vector search simultaneously.

This technique is essential for building "Smart Search" interfaces that understand dates, categories, and ranges without complex form-filling.

Learning Goals

  • Define the Self-Query pattern and its use cases.
  • Implement the SelfQueryRetriever in LangChain.
  • Map metadata field descriptions to help the LLM generate accurate filters.

Core Concepts

1. Natural Language to Structured Query

The LLM acts as a translator.

  • User: "Show me horror movies from the 90s."
  • LLM Output:
    • Query: "horror movies"
    • Filter: {"genre": "horror", "year": {"$gte": 1990, "$lt": 2000}}

2. Attribute Information

To work correctly, the Self-Query retriever needs to know what metadata fields exist and what they mean. You provide a list of AttributeInfo objects that define the names and descriptions of your columns.

3. Search + Filter (Integrated)

The retriever performs a single call to the vector database that includes both the semantic query and the Boolean filter, ensuring high efficiency and accuracy.

Self-Query Workflow

Implementing Self-Query Retrieval

  1. 1
    Step 1

    Describe your metadata fields so the LLM knows how to filter them:

    1from langchain.chains.query_constructor.base import AttributeInfo
2
3metadata_field_info = [
4    AttributeInfo(name="genre", description="The genre of the movie", type="string"),
5    AttributeInfo(name="year", description="The year the movie was released", type="integer"),
6    AttributeInfo(name="rating", description="A 1-10 rating for the movie", type="float"),
7]
8document_content_description = "Brief summary of a movie"
  2. 2
    Step 2

    Use a reasoning model like gpt-4o for the best parsing accuracy:

    1from langchain.retrievers.self_query.base import SelfQueryRetriever
2from langchain_openai import ChatOpenAI
3
4llm = ChatOpenAI(temperature=0)
5retriever = SelfQueryRetriever.from_llm(
6    llm,
7    vector_store,
8    document_content_description,
9    metadata_field_info,
10    verbose=True
11)
  3. 3
    Step 3

    The system will now handle the logic automatically:

    1# This generates a filter for rating > 8.5
2results = retriever.invoke("What are some highly rated movies about space?")

Imagine an e-commerce site. A user types: "Show me blue Nike shoes under $100."

  1. Semantic Query: "Nike shoes"
  2. Filter: {"color": "blue", "brand": "Nike", "price": {"$lt": 100}} The system retrieves the exact products requested without the user ever touching a dropdown menu or price slider.

Common Mistakes

  • Vague Descriptions: If your AttributeInfo description is just "The year," the LLM might struggle. Be specific: "The 4-digit release year of the document."
  • Unsupported Vector Stores: Not all vector stores support Self-Query filtering. Check the LangChain documentation for your specific database (Chroma, Pinecone, and Qdrant are well-supported).

Recap

  • Self-Querying turns an LLM into a sophisticated query parser.
  • It combines structured data filtering with unstructured semantic search.
  • It provides a "zero-config" search experience for end users.

Knowledge Check

Question 1 of 3
Q1Single choice

What is the primary role of the LLM in a Self-Query Retriever?

Previous

Next

Self-Query Retriever | Retrieval-Augmented Generation (RAG) — From Fundamentals to Production-Ready Agentic RAG Systems | Coursify