Self-Querying Retriever

Overview

Self-Querying Retriever uses an LLM to decompose natural language queries into two components: a semantic search query and structured metadata filters. This enables more precise retrieval than vector search alone.

The Problem

User: "Find recent articles about Python written after 2023"

Standard vector search:

• Embeds entire query
• Can't separate semantic intent from filters
• May retrieve old articles or non-Python content

How Self-Querying Works

Query Decomposition

LLM breaks query into:

1. Semantic query: "articles about Python"
2. Metadata filters: {"year": {"$gt": 2023}, "language": "Python"}

Execution

results = vectorstore.search(
    query="articles about Python",  # Semantic
    filter={"year": {"$gt": 2023}}  # Structured
)

Implementation

from langchain.retrievers.self_query.base import SelfQueryRetriever
from langchain.chains.query_constructor.base import AttributeInfo

# Define metadata schema
metadata_field_info = [
    AttributeInfo(
        name="year",
        description="The year the document was published",
        type="integer",
    ),
    AttributeInfo(
        name="language",
        description="Programming language",
        type="string",
    ),
]

retriever = SelfQueryRetriever.from_llm(
    llm=llm,
    vectorstore=vectorstore,
    document_contents="Articles about programming",
    metadata_field_info=metadata_field_info,
)

docs = retriever.get_relevant_documents(
    "Recent Python articles from 2024"
)

Benefits

• Precision: Filters irrelevant results
• Natural Language: Users don't need to know filter syntax
• Efficiency: Pre-filters before distance computation
• Better Results: Combines semantic + structured search

Example Queries

"Movies with Tom Hanks from the 1990s" → Semantic: "Tom Hanks movies" → Filter: {"year": {"$gte": 1990, "$lt": 2000}}

"Cheap hotels near the beach" → Semantic: "hotels near beach" → Filter: {"price": {"$lt": 100}, "location": "beach"}

Requirements

• Vector database with metadata filtering
• LLM for query decomposition
• Well-defined metadata schema
• Properly indexed metadata fields

Pricing

Adds small LLM API cost per query for decomposition.