updated ui tester

app.py

@@ -102,22 +102,26 @@ async def generate_endpoint(
 
 @app.post("/ui-tester")
 async def ui_tester(file: UploadFile = File(...), description: str = Query(...)):
-    """Uses BLIP's native capability to score the match between image and text."""
     image = Image.open(file.file).convert("RGB")
     blip_data = MODELS["blip"]
     
-    # We use the processor to prepare both image and text for the model
     inputs = blip_data["processor"](images=image, text=description, return_tensors="pt").to(DEVICE)
     
     with torch.no_grad():
-        # BLIP models have a built-in vision/text matching logic
-        # For simple captioning models, we can use the model's loss or log-likelihood
         outputs = blip_data["model"](**inputs, labels=inputs["input_ids"])
-        # We convert the loss to a pseudo-similarity score (lower loss = higher match)
         loss = outputs.loss.item()
-        score = 1 / (1 + loss) # Normalized 0 to 1
+        
+        # Scaling the score to make 0.3 look like a "Strong Match"
+        # and 0.2 look like a "Poor Match"
+        # Using a steep sigmoid or a linear multiplier:
+        score = max(0, min(1, (1 / (loss + 1e-6)) * 0.5)) 
+    
+    # Define thresholds based on your tests
+    status = "Match Found" if score > 0.25 else "No Match"
     
     return {
-        "match_score": round(score, 4),
-        "status": "High match" if score > 0.4 else "Low match"
+        "raw_loss": round(loss, 4),
+        "confidence_score": round(score, 4),
+        "status": status,
+        "is_valid": score > 0.25
     }