Upload app.py

app.py

@@ -886,13 +886,6 @@ planner_prompt = f"""
     - NEVER issue a final verdict without calling `execute_page_expert`
     - If no page contains sufficient proof, return **Unverified**
     - Prefer false negatives over false positives
-    *** CRITICAL VISUAL PROTOCOL ***
-    - When `execute_page_expert` returns, it will explicitly state "VISUAL_PROOF_PENDING".
-    - When you see this, your ONLY response must be: "Awaiting visual proof."
-    - DO NOT attempt to guess the verdict.
-    - DO NOT complain about missing images.
-    - Simply wait. The user will immediately send the images in the next turn.
-    
     
     ========================
     QUALITY STANDARD
@@ -912,91 +905,83 @@ chat = planner.chats.create(model=planner_model, config=config)
 def agent_worker(user_question):
     state.clear()
     state.add_log(f'🚀 Starting analysis for: **{user_question}**')
+    state.add_analysis("🧠 Planner initialized. Awaiting tool calls...")
 
-    # 1. Start the conversation
     response = chat.send_message(user_question)
     
-    # 2. Loop through tool calls
     while response.candidates[0].content.parts[0].function_call:
         tool_responses = []
-        pending_images = [] # Reset every turn
+        pending_images = []
 
         for part in response.candidates[0].content.parts:
             if part.function_call:
                 name = part.function_call.name
                 args = part.function_call.args
+
                 state.add_log(f'🛠️ Planner calling: **{name}**')
+                state.add_analysis(
+                    f"### 🛠️ Tool Call: `{name}`\n"
+                    f"```json\n{json.dumps(args, indent=2)}\n```"
+                )
 
-                # Execute tool
                 func = globals()[name]
                 result = func(**args)
 
-                # 1. Handle Visuals
+                # -----------------------------
+                # STREAM REAL TOOL OUTPUTS
+                # -----------------------------
+
+                # search_page_text
+                # execute_page_expert
+                
                 if name == "execute_page_expert":
+
                     tile_idxs = result.get("visual_pointers", [])
                     page_num = args.get("page_num")
 
                     if tile_idxs:
-                        state.add_log(f'📸 Staging images for tiles: {tile_idxs}')
-                        stitched_bytes = merge_tiles(tile_idxs, page_num)
+                        state.add_log(f'📸 Stitching high-res proof for tiles: **{tile_idxs}**')
+                        state.add_analysis(
+                            f"📸 Visual proof requested for tiles `{tile_idxs}` on page `{page_num}`"
+                        )
+
+                        stitched_bytes = merge_tiles(
+                            tile_indexes=tile_idxs,
+                            page_num=page_num
+                        )
                         
-                        # Add to pending images
-                        pending_images.append(types.Part.from_bytes(stitched_bytes, mime_type="image/png"))
-                        pending_images.append(types.Part.from_bytes(image_bytes_list[page_num], mime_type="image/png"))
+                        pending_images.append(
+                            types.Part.from_bytes(stitched_bytes, mime_type="image/png")
+                            )
+                        pending_images.append(types.Part.from_bytes(
+                                image_bytes_list[page_num],
+                                mime_type="image/png"
+                            ))
                         
-                        # Update text result to reference the incoming images
-                        result["note"] = "Visual evidence generated. See next message for images."
 
-                # 2. Collect the Tool Response Part
                 tool_responses.append(
                     types.Part.from_function_response(
                         name=name,
                         response={"result": result}
                     )
                 )
+                
 
-        state.add_analysis("🧠 Injecting tool outputs and sending images...")
-
-        # =========================================================================
-        # THE GPT-STYLE FIX: Manual History Injection
-        # =========================================================================
-        
+        state.add_analysis("🧠 Returning tool outputs to planner...")
+        response = chat.send_message(tool_responses)
         if pending_images:
-            # Step A: Manually append the Tool Responses to history.
-            # We create a Content object (or dict) with role='function'.
-            # This "closes" the function loop in the history without triggering the model yet.
-            
-            # Note: Depending on your specific SDK version, you might need 
-            # from google.ai.generativelanguage_v1beta.types import Content
-            # But usually a dict works fine in the python SDK:
-            
-            tool_content = {
-                "role": "function",
-                "parts": tool_responses
-            }
-            chat.history.append(tool_content)
-
-            # Step B: Send the images as the "User" follow-up.
-            # The model sees: [FunctionCall] -> [FunctionResponse] -> [User Images]
-            # It will now generate the Verdict based on both.
-            
-            image_message = [
-                types.Part.from_text(
-                    "Here is the VISUAL PROOF generated by the tool. "
-                    "Analyze these images to confirm the compliance verdict."
-                )
-            ] + pending_images
+            state.add_log(f'📸 Sending {len(pending_images)} images to Planner...')
             
-            response = chat.send_message(image_message)
+            # We send the images with a prompt telling the model what they are
+            pending_images.insert(0, "Here is the visual proof generated by the tool. Please use this to confirm your final answer:")
             
-        else:
-            # Standard path: If no images, just send the tool response normally
-            response = chat.send_message(tool_responses)
+            # This generates the ACTUAL final answer that sees the image
+            response = chat.send_message(pending_images)
 
     state.add_log('🏁 **ANALYSIS COMPLETE**')
+    state.add_analysis("✅ Planner finished. Final verdict generated.")
     state.final_answer = response.text
     state.done = True
-
     
 def run_agentic_workflow(user_question):
     state.done = False