Metadata-Based Filtering

Metadata filtering is the most effective way to eliminate "irrelevant context" from your RAG pipeline before it even hits the LLM.

Pre-Filtering vs. Post-Filtering

Post-Filtering (Inefficient)

1. Search the whole vector DB for Top 10 documents.
2. Filter the result list to keep only "Admin" documents.
3. Problem: You might end up with 0 documents if none of the Top 10 matched the filter.

Pre-Filtering (Chroma's Way)

1. Find all documents where access_role == 'Admin'.
2. Perform vector search only on those documents.
3. Benefit: You always get the best matches that actually meet your constraints.

Advanced Filter Logic in RAG

Temporal Decay

Often, you want information that is both relevant (vector similarity) and fresh (recent date).

# Search docs from the last 90 days
three_months_ago = "2023-10-01" 
results = collection.query(
    query_texts=["quarterly results"],
    where={"date": {"$gte": three_months_ago}}
)

Geographical/Regional Context

In global RAG systems, you might want to filter by region: where={"region": "EMEA"}

Citation-Quality Filtering

You might only want to retrieve chunks that have a high "Source Quality" score or those that have been "Verified" by a human.

Implementation Tip: Flat Metadata

Most vector databases perform better with "Flat" metadata (key-value pairs) rather than nested JSON.

• Bad: metadata={"info": {"author": "Sudeep", "date": "2024"}}
• Good: metadata={"author": "Sudeep", "date": "2024"}

Exercises

1. Implement a search that only returns "High Confidence" documents (using a confidence field in your metadata).
2. What happens if you apply a filter that is too restrictive (e.g., searching for text that doesn't exist in a specific department)?
3. How does metadata filtering help prevent "Data Leakage" between users?