Medical GraphRAG Assistant

# Memory Search Browse Fix - Empty Query Support ## Issue Memory search was failing with error: `'<' not supported between instances of 'str' and 'float'` when trying to search. Additionally, empty search strings should allow browsing all memories, but instead showed a warning. ## Root Causes ### 1. Type Comparison Error **Location**: `src/memory/vector_memory.py:207` **Problem**: InterSystems IRIS was returning similarity scores as strings instead of floats, causing type error when comparing `similarity < min_similarity`. **Fix**: Added explicit type conversion with error handling: ```python # Convert similarity to float (IRIS may return as string) try: similarity = float(similarity) if similarity is not None else 0.0 except (ValueError, TypeError): similarity = 0.0 ``` ### 2. Empty Query Not Supported **Location**: `src/memory/vector_memory.py:156` (recall method) **Problem**: Empty search strings triggered warning "Please enter a search query" instead of allowing users to browse all memories. **Fix**: Added special handling for empty queries - returns all memories sorted by use count and recency: ```python # Handle empty query - return all memories sorted by use count if not query or not query.strip(): # ... SQL query without vector search ORDER BY UseCount DESC, UpdatedAt DESC ``` ### 3. UI Warning for Empty Search **Location**: `mcp-server/streamlit_app.py:988-991` **Problem**: Streamlit UI showed warning and blocked empty searches. **Fix**: - Removed the warning check - Pass empty string to recall() to trigger browse mode - Updated placeholder text to indicate empty search is supported ## Changes Made ### `src/memory/vector_memory.py` 1. **Lines 169-211**: Added empty query handler that returns all memories sorted by use_count 2. **Lines 246-250**: Added type conversion for similarity scores with try/except ### `mcp-server/streamlit_app.py` 1. **Line 982**: Updated placeholder text: `"e.g., 'pneumonia' or leave empty to browse all"` 2. **Lines 988-991**: Simplified search button logic to allow empty queries ## Features Added ### Browse All Memories - Click search button with empty query → shows all memories - Results sorted by use count (most used first) - Then by last updated timestamp - Similarity score shows as 1.0 (no vector search performed) ### Semantic Search - Enter search query → uses NV-CLIP embeddings for semantic similarity - Results sorted by cosine similarity - Shows actual similarity scores (0-1) - Filters by min_similarity threshold (default 0.5) ## Testing ### Test Empty Search (Browse Mode) ```python from src.memory import VectorMemory memory = VectorMemory() # Empty query returns all memories sorted by use count results = memory.recall("", top_k=10) for r in results: print(f"{r['memory_type']}: {r['text'][:50]}... (used {r['use_count']}x)") ``` ### Test Semantic Search ```python # Search with query uses vector similarity results = memory.recall("pneumonia", top_k=5) for r in results: print(f"Similarity {r['similarity']:.2f}: {r['text'][:50]}...") ``` ## Current Status - ✅ Streamlit restarted with fixes applied - ✅ SSH tunnel to AWS NIM active on port 8002 - ✅ Empty search works - browses all memories - ✅ Semantic search works - uses NV-CLIP embeddings - ✅ Type errors fixed - handles string/float conversions - ✅ UI updated with clear placeholder text ## Usage ### Browse All Memories 1. Open http://localhost:8501 2. Expand sidebar "📚 Browse Memories" 3. Leave search box empty 4. Click 🔍 Search 5. See all 5 memories sorted by use count ### Search Memories Semantically 1. Enter search query (e.g., "chest", "pneumonia", "diagnosis") 2. Click 🔍 Search 3. See results ranked by semantic similarity ## Files Modified - `src/memory/vector_memory.py` - Added empty query handling and type conversion - `mcp-server/streamlit_app.py` - Removed empty query warning, updated placeholder ## Related Fixes - See `EMBEDDINGS_FIXED.md` for the original embedding issues - This fix completes the memory search feature implementation