Replace generic 'Retrieved N documents' message with actual answer synthesis
from the top 3 documents when LLM synthesis fails. Extracts:
- Document previews (first 300 chars)
- Document filenames from metadata
- Creates readable answer format
This ensures users get useful information even when LLM synthesis encounters errors.
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
Wrap synthesizer.synthesize() in try-catch to handle AttributeError gracefully
and fall back to returning raw documents when LLM synthesis fails.
This allows the chat feature to work even when there are Pydantic or llama-index
compatibility issues with TextNode attributes.
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
The SimilarityPostprocessor tries to access node.score which isn't supported
in the current llama-index version and causes Pydantic validation errors.
Since we're already getting sorted results from vector similarity search,
disable the postprocessor and use retrieved nodes directly.
This fixes chat and search endpoints.
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
TextNode doesn't support a direct score attribute in this version of llama-index.
Instead, store the similarity score in the node's metadata during creation and
retrieve it from metadata during response formatting.
This ensures full compatibility with Pydantic validation and prevents attribute errors.
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
Instead of setting node.score after creation (which fails Pydantic validation),
pass the score parameter during TextNode initialization. This ensures the score
field is properly validated during object creation.
This fixes the error: 'Object has no attribute score' in chat and search endpoints.
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
The similarity score was being set as a numpy.float32 from SQLite distance calculations.
Pydantic expects Python native float types for the score field, causing validation errors
when querying MCPs or using chat features.
Fix: Convert similarity score to Python float using float() conversion.
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
Features:
- Fixed Ollama configuration for native (localhost) instead of Docker
- Increased LLM timeout from 30s to 120s for quality answers
- Implemented KB slug system for clean MCP endpoint URLs
- Added multiple knowledge base support with type metadata
- Integrated MCP server with FastAPI (port 8051)
- React frontend with KB management UI
- PostgreSQL + pgvector for vector embeddings
Performance:
- 33-56% faster than Docker setup with native Ollama
- Direct Metal GPU acceleration on Apple Silicon
- Optimized RAG engine with caching
- ~40s response time for comprehensive answers
Documentation:
- Complete native setup guide (README_NATIVE_SETUP.md)
- Performance benchmarking and comparison
- MCP endpoint documentation
- Slug implementation details
- Decision rationale for native vs Docker
Cleanup:
- Removed old Streamlit setup files
- Consolidated setup scripts to single setup_native.sh
- Consolidated service management scripts
- Removed test artifacts and old analysis docs
- Ready for sharing with team
🚀 Production-ready for local team deployment
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
