Update backend/app/openrouter_client.py

backend/app/openrouter_client.py

@@ -15,9 +15,15 @@ MODEL_NAME = "qwen/qwen3-vl-235b-a22b-instruct"
 def _file_to_image_block(file_bytes: bytes, content_type: str) -> Dict[str, Any]:
     """
     Encode the file as a data URL to feed into the multimodal model.
-    For demo purposes we treat PDFs and images the same way here.
+    Note: PDFs may not be directly supported by vision models.
+    For images (PNG, JPG, etc.), this works fine.
+    For PDFs, the model might not be able to process them.
     """
     b64 = base64.b64encode(file_bytes).decode("utf-8")
+    
+    # Log file type for debugging
+    print(f"[DEBUG] Encoding file as image block. Content type: {content_type}, Size: {len(file_bytes)} bytes")
+    
     return {
         "type": "input_image",
         "image_url": f"data:{content_type};base64,{b64}",
@@ -101,7 +107,13 @@ async def extract_fields_from_document(
         data = resp.json()
 
     # OpenRouter returns choices[0].message.content
+    if "choices" not in data or len(data["choices"]) == 0:
+        raise ValueError("No choices in OpenRouter response")
+    
     content = data["choices"][0]["message"]["content"]
+    
+    # Log the raw response for debugging (first 500 chars)
+    print(f"[DEBUG] OpenRouter response preview: {str(content)[:500]}")
 
     # content may be a string or a list of content blocks
     if isinstance(content, list):
@@ -109,34 +121,46 @@ async def extract_fields_from_document(
     else:
         text = content
 
+    if not text or not text.strip():
+        raise ValueError("Empty response from OpenRouter API")
+
     # Try to parse JSON from the model output
     # The model might return JSON wrapped in markdown code blocks or with extra text
     try:
         # First, try direct JSON parsing
-        return json.loads(text)
-    except json.JSONDecodeError:
+        parsed = json.loads(text)
+        print(f"[DEBUG] Successfully parsed JSON directly")
+        return parsed
+    except json.JSONDecodeError as e:
+        print(f"[DEBUG] Direct JSON parse failed: {e}")
         # Try to extract JSON from markdown code blocks
         json_match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', text, re.DOTALL)
         if json_match:
             try:
-                return json.loads(json_match.group(1))
-            except json.JSONDecodeError:
-                pass
+                parsed = json.loads(json_match.group(1))
+                print(f"[DEBUG] Successfully parsed JSON from markdown code block")
+                return parsed
+            except json.JSONDecodeError as e2:
+                print(f"[DEBUG] Markdown code block parse failed: {e2}")
         
         # Try to find JSON object in the text (look for {...})
         json_match = re.search(r'\{.*\}', text, re.DOTALL)
         if json_match:
             try:
-                return json.loads(json_match.group(0))
-            except json.JSONDecodeError:
-                pass
+                parsed = json.loads(json_match.group(0))
+                print(f"[DEBUG] Successfully parsed JSON from regex match")
+                return parsed
+            except json.JSONDecodeError as e3:
+                print(f"[DEBUG] Regex match parse failed: {e3}")
         
         # If all parsing fails, return a default structure with the raw text
+        print(f"[WARNING] All JSON parsing attempts failed. Returning fallback structure.")
         return {
             "doc_type": "other",
             "confidence": 50.0,
             "fields": {
-                "raw_text": text[:500],  # First 500 chars
-                "error": "Could not parse JSON from model response"
+                "raw_response": text[:1000],  # First 1000 chars for debugging
+                "error": "Could not parse JSON from model response",
+                "note": "Check server logs for full response"
             }
         }