Medical GraphRAG Assistant

#!/usr/bin/env python3 """ Fix Memory Embeddings - Re-embed existing memories with real NV-CLIP vectors This script re-generates embeddings for all memories that have mock/zero embeddings. Requires NVCLIP_BASE_URL to be set to a working NV-CLIP endpoint. """ import sys import os # Add parent directory to path parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), '..')) if parent_dir not in sys.path: sys.path.insert(0, parent_dir) from src.db.connection import get_connection from src.embeddings.nvclip_embeddings import NVCLIPEmbeddings def fix_memory_embeddings(): """Re-embed memories with real NV-CLIP vectors.""" print("=" * 60) print("Fix Memory Embeddings with NV-CLIP") print("=" * 60) print() # Initialize NV-CLIP try: embedder = NVCLIPEmbeddings() print(f"✅ NV-CLIP initialized") print() except Exception as e: print(f"❌ Failed to initialize NV-CLIP: {e}") print() print("Make sure NVCLIP_BASE_URL is set:") print(" export NVCLIP_BASE_URL=http://localhost:8002/v1") return 1 # Connect to database conn = get_connection() cursor = conn.cursor() try: # Get all memories cursor.execute(""" SELECT MemoryID, MemoryText, VECTOR_DOT_PRODUCT(Embedding, Embedding) as magnitude FROM SQLUser.AgentMemoryVectors """) memories = cursor.fetchall() total = len(memories) if total == 0: print("No memories found in database.") return 0 print(f"Found {total} memories") print() # Count how many need fixing needs_fix = sum(1 for mem in memories if mem[2] == 0.0) print(f"Memories with mock embeddings (zeros): {needs_fix}") print(f"Memories with real embeddings: {total - needs_fix}") print() if needs_fix == 0: print("✅ All memories already have real embeddings!") return 0 # Fix memories print(f"Re-embedding {needs_fix} memories...") print() fixed = 0 errors = 0 for memory_id, memory_text, magnitude in memories: if magnitude > 0: continue # Skip memories that already have real embeddings try: # Generate real embedding embedding = embedder.embed_text(memory_text) # Verify it's real import numpy as np vec = np.array(embedding) new_magnitude = np.dot(vec, vec) if new_magnitude == 0: print(f" ❌ {memory_id[:8]}: Still got zeros") errors += 1 continue # Update in database cursor.execute(""" UPDATE SQLUser.AgentMemoryVectors SET Embedding = TO_VECTOR(?) WHERE MemoryID = ? """, (str(embedding), memory_id)) conn.commit() fixed += 1 print(f" ✅ {memory_id[:8]}: Fixed (magnitude {new_magnitude:.4f})") except Exception as e: print(f" ❌ {memory_id[:8]}: Error - {e}") errors += 1 print() print("=" * 60) print("Summary") print("=" * 60) print(f"Total memories: {total}") print(f"Fixed: {fixed}") print(f"Errors: {errors}") print(f"Already had real embeddings: {total - needs_fix}") print() if fixed > 0: print("✅ Memory search should now work!") return 0 finally: cursor.close() conn.close() if __name__ == '__main__': sys.exit(fix_memory_embeddings())