Add dynamic metadata filtering for RAG queries

[samkeen]

Feature Request: Dynamic Metadata Filtering for RAG Queries

Problem Statement

Currently, the RAG implementation only supports filtering by:

• Number of results (n_results)
• Distance threshold (distance_threshold)

However, ChromaDB supports powerful metadata filtering through its where parameter, which could significantly improve the precision of document retrieval.

Proposed Solution

Add a dynamic filter builder UI that allows users to create metadata-based filters for their RAG queries.

Implementation Overview

Backend Changes

1. RAG Config Service (rag_config_service.py)

   • Add method to detect available metadata fields from collection
   • Modify query_collection() to accept where parameter
   • Store filter preferences in config

2. API Endpoints (routes.py)

   • GET /api/rag/metadata-fields - Return available fields with types and unique values
   • Update POST /api/chat to accept filter parameters

Frontend Changes

1. Main Chat Interface (script.js, index.html)

   • Collapsible filter panel below RAG toggle
   • Dynamic filter rows with field/operator/value selectors
   • Support AND/OR logic between filters
   • Show active filter count badge

2. Settings Page (settings.js, settings.html)

   • Preview available metadata fields when collection selected
   • Configure default filters

Filter Types Support

• Text fields: equals, contains (using $in)
• Numbers: equals, $gt, $lt, $gte, $lte, range
• Lists: multi-select with $in/$nin
• Dates: date picker with comparison operators

Example Filter Format

{
  "filters": [
    {"field": "author", "operator": "$eq", "value": "John Doe"},
    {"field": "chapter", "operator": "$in", "value": [1, 2, 3]},
    {"field": "date", "operator": "$gte", "value": "2024-01-01"}
  ],
  "logic": "$and"  // or "$or"
}

User Benefits

1. Precision: Target specific document subsets (e.g., "only search in chapter 3")
2. Efficiency: Reduce noise from irrelevant content
3. Flexibility: Build complex queries without writing code
4. Discovery: Explore metadata patterns in the corpus
5. Performance: Smaller, more relevant result sets

Additional Features to Consider

• Save/load filter presets
• Quick filter templates ("Recent docs", "By author")
• Filter match explanations in results
• Visual indicators for active filters
• Recently used filters history

ChromaDB Reference

ChromaDB supports these metadata filter operators:

• Comparison: $eq, $ne, $gt, $gte, $lt, $lte
• Logical: $and, $or
• Inclusion: $in, $nin

Documentation: https://docs.trychroma.com/docs/querying-collections/metadata-filtering

Acceptance Criteria

• Users can add/remove filter conditions dynamically
• Filters persist across page refreshes
• Filter UI shows available fields from current collection
• Applied filters are visible in chat details modal
• Clear documentation on how to use filters