Update app.py

app.py

@@ -20,53 +20,40 @@ def load_data():
     print(f"Number of patent numbers: {len(patent_numbers)}")
     print(f"Number of metadata entries: {len(metadata)}")
     
-    # Print sample metadata
-    sample_patent = next(iter(metadata))
-    print(f"Sample metadata for patent {sample_patent}:")
-    print(json.dumps(metadata[sample_patent], indent=2))
-    
     return embeddings, patent_numbers, metadata
 
 embeddings, patent_numbers, metadata = load_data()
 
-# Create FAISS index
-index = faiss.IndexFlatL2(embeddings.shape[1])
+# Normalize embeddings for cosine similarity
+embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)
+
+# Create FAISS index for cosine similarity
+index = faiss.IndexFlatIP(embeddings.shape[1])
 index.add(embeddings)
 
 # Load BERT model for encoding search queries
-embedding_dim = embeddings.shape[1]
-print(f"Embedding dimension: {embedding_dim}")
+model = SentenceTransformer('all-mpnet-base-v2')
 
-if embedding_dim == 384:
-    model = SentenceTransformer('all-MiniLM-L6-v2')
-elif embedding_dim == 768:
-    model = SentenceTransformer('all-mpnet-base-v2')
-else:
-    print(f"Unexpected embedding dimension: {embedding_dim}")
-    model = SentenceTransformer('all-MiniLM-L6-v2')  # Default to this model
+def exact_text_match(query, metadata):
+    matches = []
+    for patent_number, data in metadata.items():
+        if query.lower() in data.get('abstract', '').lower() or query.lower() in data.get('claims', '').lower():
+            matches.append((patent_number, 1.0))  # Score of 1.0 for exact match
+    return matches
 
 def search(query, top_k=5):
+    print(f"Searching for: {query}")
+    
     # Encode the query
     query_embedding = model.encode([query])[0]
+    query_embedding = query_embedding / np.linalg.norm(query_embedding)
     
-    # Debug: Print query embedding
-    print(f"Query embedding: {query_embedding}")
-    
-    # Ensure the query embedding has the same dimension as the index
-    if query_embedding.shape[0] != index.d:
-        print(f"Query embedding dimension ({query_embedding.shape[0]}) does not match index dimension ({index.d})")
-        # Option 1: Pad or truncate the query embedding
-        if query_embedding.shape[0] < index.d:
-            query_embedding = np.pad(query_embedding, (0, index.d - query_embedding.shape[0]))
-        else:
-            query_embedding = query_embedding[:index.d]
+    print(f"Query embedding shape: {query_embedding.shape}")
     
     # Perform similarity search
     distances, indices = index.search(np.array([query_embedding]), top_k)
     
-    # Debug: Print distances and indices
-    print(f"Distances: {distances}")
-    print(f"Indices: {indices}")
+    print(f"FAISS search results - Distances: {distances}, Indices: {indices}")
     
     results = []
     for i, idx in enumerate(indices[0]):
@@ -76,17 +63,23 @@ def search(query, top_k=5):
             continue
         patent_data = metadata[patent_number]
         result = f"Patent Number: {patent_number}\n"
-        
-        # Safely extract abstract
         abstract = patent_data.get('abstract', 'No abstract available')
-        if isinstance(abstract, str):
-            result += f"Abstract: {abstract[:200]}...\n"
-        else:
-            result += f"Abstract: Unable to display (type: {type(abstract)})\n"
-        
-        result += f"Similarity Score: {1 - distances[0][i]:.4f}\n\n"
+        result += f"Abstract: {abstract[:200]}...\n"
+        result += f"Similarity Score: {distances[0][i]:.4f}\n\n"
         results.append(result)
     
+    # Fallback to exact text match if no results or low similarity
+    if not results or distances[0][0] < 0.5:
+        print("Falling back to exact text match")
+        exact_matches = exact_text_match(query, metadata)
+        for patent_number, score in exact_matches[:top_k]:
+            patent_data = metadata[patent_number]
+            result = f"Patent Number: {patent_number}\n"
+            abstract = patent_data.get('abstract', 'No abstract available')
+            result += f"Abstract: {abstract[:200]}...\n"
+            result += f"Exact Match Score: {score:.4f}\n\n"
+            results.append(result)
+    
     return "\n".join(results)
 
 # Create Gradio interface