Improve Visual Search with Negative Vectors

app.py

@@ -12,6 +12,7 @@ from PIL import Image, ImageDraw, ImageFont
 import io
 import zipfile
 from huggingface_hub import hf_hub_download
+import numpy as np
 
 app = FastAPI()
 
@@ -134,7 +135,7 @@ async def home(request: Request):
 async def search(request: Request, q: str, searchmode: str = "text"):
     if not q: return ""
 
-    # 1. ANALYTICS
+    # 1. ANALYTICS (Keep existing code)
     try:
         conn = get_db_connection()
         c = conn.cursor()
@@ -151,19 +152,39 @@ async def search(request: Request, q: str, searchmode: str = "text"):
     results = []
     seen_files = set()
 
-    # --- MODE 1: VISUAL SEARCH ---
+    # --- MODE 1: VISUAL SEARCH (UPGRADED LOGIC) ---
     if searchmode == "visual" and visual_tbl:
         try:
-            # Encode text query to Visual Vector Space
-            query_vec = visual_model.encode(q)
-            vec_results = visual_tbl.search(query_vec).limit(20).to_list()
+            # A. ENHANCE THE POSITIVE (Prompt Engineering)
+            # CLIP works better when you tell it "This is a photo"
+            positive_prompt = f"a photo of {q}, visual evidence"
+            pos_vec = visual_model.encode(positive_prompt)
+
+            # B. DEFINE THE NEGATIVE (The "Noise" to remove)
+            # We tell the AI: "Ignore boring text pages"
+            negative_prompt = "scanned text document, white paper, handwriting, typed text"
+            neg_vec = visual_model.encode(negative_prompt)
+
+            # C. VECTOR MATH (The Magic)
+            # Formula: Target = UserQuery - (0.5 * BoringPaper)
+            # This pushes the search vector AWAY from text and TOWARDS the object
+            alpha = 0.5 
+            final_vec = pos_vec - (alpha * neg_vec)
+
+            # Normalize vector (Important for accurate cosine similarity)
+            norm = np.linalg.norm(final_vec)
+            if norm > 0:
+                final_vec = final_vec / norm
+
+            # Search with the new "Clean" vector
+            vec_results = visual_tbl.search(final_vec).limit(20).to_list()
             
             for res in vec_results:
                 results.append({
                     "type": "Visual Match",
                     "filename": res['filename'],
                     "page": res['page'],
-                    "text": f"Image match for '{q}'",
+                    "text": f"Visual match for '{q}'",
                     "score": 1.0 - res['_distance']
                 })
         except Exception as e:
@@ -171,8 +192,8 @@ async def search(request: Request, q: str, searchmode: str = "text"):
             
         return templates.TemplateResponse("partials/results.html", {"request": request, "results": results})
 
-    # --- MODE 2: TEXT/HYBRID SEARCH ---
-    
+    # --- MODE 2: TEXT SEARCH (Keep existing code exactly as is) ---
+    # ... (Paste your existing text search logic here) ...
     # A. SQLite Keyword Search
     conn = get_db_connection()
     cursor = conn.cursor()