Update app.py

app.py

@@ -40,7 +40,7 @@ class Config(object):
     def __init__(self):
         self.random_state = 42
         self.max_len = 256
-        self.reasoning_max_len = 128
+        self.reasoning_max_len = 256
         self.temperature = 0.1
         self.repetition_penalty = 1.2
         self.DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
@@ -173,6 +173,7 @@ class AgentState(TypedDict):
     confidence: float
     judge_explanation: str
 
+
 ALL_TOOLS = {
     "web_search": ["query"],
     "visit_webpage": ["url"],
@@ -363,7 +364,7 @@ def visit_webpage(url: str) -> str:
     if content is not None:
         for table in content.find_all("table", {"class": "wikitable"}):
             for row in table.find_all("tr"):
-                cols = [c.get_text(strip=True) for c in row.find_all(["td", "th"])]
+                cols = [c.get_text(strip=False) for c in row.find_all(["td", "th"])]
                 if cols:
                     table_texts.append(" | ".join(cols))
 
@@ -512,12 +513,12 @@ Response: <answer>
 
 DO NOT add anything additional and return ONLY what is asked and in the format asked.
 
-If you output anything else, it is incorrect.
-
 ONLY return a response if you are confident about the answer, otherwise return empty string.
 
 If you output anything else, it is incorrect.
 
+If there is no information provided or the information is not relevant then answer as best based on your own knowledge.
+
 Example of valid json response for user request: Who was the winner of 2025 World Snooker Championship:
 Response: Zhao Xintong.
 
@@ -531,8 +532,8 @@ Information:
 {information}
 """
 
-    # raw_output = reasoning_generate(prompt)
-    raw_output = generate(prompt)
+    raw_output = reasoning_generate(prompt)
+    # raw_output = generate(prompt)
 
     logger.info(f"Raw Output: {raw_output}")
 
@@ -550,16 +551,32 @@ Information:
 
 
     raw = raw_output.strip()
+
+    matches = re.findall(r"Response:\s*([^\n]+)", raw)
     
-    # Find the first valid "Response: ..." occurrence
-    match = re.search(r"Response:\s*([^\n\.]+)", raw)
-    
-    if match:
-        output = match.group(1).strip()
+    if matches:
+        output = matches[-1].strip()  # ✅ take LAST occurrence
     else:
-        # fallback: take first line
-        output = raw.split("\n")[0].strip()
-
+        # Find the first valid "Response: ..." occurrence
+        match = re.search(r"Response:\s*([^\n\.]+)", raw)
+        
+        if match:
+            output = match.group(1).strip()
+        else:
+            # fallback: take first line
+            output = raw.split("\n")[0].strip()
+    
+        if "Response:" in output:
+            output = output.split("Response:")[-1]
+        elif "Response" in output:
+            output = output.split("Response")[-1]
+        
+        # Clean quotes / trailing punctuation
+        output = output.strip('"').strip()
+        if output.endswith("."):
+            output = output[:-1]
+    
+    # Clean
     if "Response:" in output:
         output = output.split("Response:")[-1]
     elif "Response" in output:
@@ -569,7 +586,31 @@ Information:
     output = output.strip('"').strip()
     if output.endswith("."):
         output = output[:-1]
+
+
+    if output == "":
+        # Find the first valid "Response: ..." occurrence
+        match = re.search(r"Response:\s*([^\n\.]+)", raw)
+        
+        if match:
+            output = match.group(1).strip()
+        else:
+            # fallback: take first line
+            output = raw.split("\n")[0].strip()
     
+        if "Response:" in output:
+            output = output.split("Response:")[-1]
+        elif "Response" in output:
+            output = output.split("Response")[-1]
+        
+        # Clean quotes / trailing punctuation
+        output = output.strip('"').strip()
+        if output.endswith("."):
+            output = output[:-1]
+
+
+    output = output.split(".")[0]
+
 
     state["output"] = output
 
@@ -655,7 +696,7 @@ Answer:
     state["confidence"] = data["confidence"]
     state["judge_explanation"] = data["explanation"]
 
-    logger.info(f"State (Judge Agent): {state}")
+    # logger.info(f"State (Judge Agent): {state}")
 
     return state
 
@@ -725,34 +766,59 @@ def tool_executor(state: AgentState):
     responsible for translating structured LLM intent into real system actions.
     """
 
-    webpage_result = ""
-    action = Action.model_validate(state["proposed_action"])
-
-    best_query_webpage_information_similarity_score = -1.0
-    best_webpage_information = ""
-
-    webpage_information_complete = ""
-
-    if action.tool == "web_search":
-        logger.info(f"action.tool: {action.tool}")
+    try:
+        webpage_result = ""
+        action = Action.model_validate(state["proposed_action"])
+    
+        best_query_webpage_information_similarity_score = -1.0
+        best_webpage_information = ""
+    
+        webpage_information_complete = ""
+    
+        if action.tool == "web_search":
+            logger.info(f"action.tool: {action.tool}")
+            
+            query_embeddings = sentence_transformer_model.encode_query(state["messages"][-1].content).reshape(1, -1)
+            query_arg_embeddings = sentence_transformer_model.encode_query(state["proposed_action"]["args"]["query"]).reshape(1, -1)
+            score = float(cosine_similarity(query_embeddings, query_arg_embeddings)[0][0])
+    
+            if score > 0.80:
+                results = web_search(**action.args)
+            else:
+                logger.info(f"Overwriting user query because the Agent suggested query had score: {state["proposed_action"]["args"]["query"]} - {score}")
+                results = web_search(**{"query": state["messages"][-1].content})
+    
+            logger.info(f"Webpages - Results: {results}")
+    
+            for result in results:
+                try:
+                    webpage_results = visit_webpage(result)
+                    webpage_result = " \n ".join(webpage_results)
+    
+                    # for webpage_result in webpage_results:
+                    query_embeddings = sentence_transformer_model.encode_query(state["messages"][-1].content).reshape(1, -1)
+                    webpage_information_embeddings = sentence_transformer_model.encode_query(webpage_result).reshape(1, -1)
+                    query_webpage_information_similarity_score = float(cosine_similarity(query_embeddings, webpage_information_embeddings)[0][0])
         
-        query_embeddings = sentence_transformer_model.encode_query(state["messages"][-1].content).reshape(1, -1)
-        query_arg_embeddings = sentence_transformer_model.encode_query(state["proposed_action"]["args"]["query"]).reshape(1, -1)
-        score = float(cosine_similarity(query_embeddings, query_arg_embeddings)[0][0])
-
-        if score > 0.80:
-            results = web_search(**action.args)
-        else:
-            logger.info(f"Overwriting user query because the Agent suggested query had score: {state["proposed_action"]["args"]["query"]} - {score}")
-            results = web_search(**{"query": state["messages"][-1].content})
-
-        logger.info(f"Webpages - Results: {results}")
-
-        for result in results:
+                    # logger.info(f"Webpage Information and Similarity Score: {result} - {webpage_result} - {query_webpage_information_similarity_score}")
+        
+                    if query_webpage_information_similarity_score > 0.65:
+                        webpage_information_complete += webpage_result
+                        webpage_information_complete += " \n "
+                        webpage_information_complete += " \n "
+        
+                    if query_webpage_information_similarity_score > best_query_webpage_information_similarity_score:
+                        best_query_webpage_information_similarity_score = query_webpage_information_similarity_score
+                        best_webpage_information = webpage_result
+    
+                except Exception as e:
+                    logger.info(f"Tool Executor - Exception: {e}")
+                    
+        elif action.tool == "visit_webpage":
             try:
-                webpage_results = visit_webpage(result)
+                webpage_results = visit_webpage(**action.args)
                 webpage_result = " \n ".join(webpage_results)
-
+    
                 # for webpage_result in webpage_results:
                 query_embeddings = sentence_transformer_model.encode_query(state["messages"][-1].content).reshape(1, -1)
                 webpage_information_embeddings = sentence_transformer_model.encode_query(webpage_result).reshape(1, -1)
@@ -768,43 +834,22 @@ def tool_executor(state: AgentState):
                 if query_webpage_information_similarity_score > best_query_webpage_information_similarity_score:
                     best_query_webpage_information_similarity_score = query_webpage_information_similarity_score
                     best_webpage_information = webpage_result
-
-            except Exception as e:
-                logger.info(f"Tool Executor - Exception: {e}")
-                
-    elif action.tool == "visit_webpage":
-        try:
-            webpage_results = visit_webpage(**action.args)
-            webpage_result = " \n ".join(webpage_results)
-
-            # for webpage_result in webpage_results:
-            query_embeddings = sentence_transformer_model.encode_query(state["messages"][-1].content).reshape(1, -1)
-            webpage_information_embeddings = sentence_transformer_model.encode_query(webpage_result).reshape(1, -1)
-            query_webpage_information_similarity_score = float(cosine_similarity(query_embeddings, webpage_information_embeddings)[0][0])
-
-            # logger.info(f"Webpage Information and Similarity Score: {result} - {webpage_result} - {query_webpage_information_similarity_score}")
-
-            if query_webpage_information_similarity_score > 0.65:
-                webpage_information_complete += webpage_result
-                webpage_information_complete += " \n "
-                webpage_information_complete += " \n "
-
-            if query_webpage_information_similarity_score > best_query_webpage_information_similarity_score:
-                best_query_webpage_information_similarity_score = query_webpage_information_similarity_score
-                best_webpage_information = webpage_result
-        except:
-            pass
-    else:
-        result = "Unknown tool"
-
-    if webpage_information_complete == "":
-        webpage_information_complete = best_webpage_information
+            except:
+                pass
+        else:
+            result = "Unknown tool"
     
-    state["information"] = webpage_information_complete[:3000]
-    state["best_query_webpage_information_similarity_score"] = best_query_webpage_information_similarity_score
-
-    logger.info(f"Information: {state['information']}")
-    logger.info(f"Information: {state['best_query_webpage_information_similarity_score']}")
+        if webpage_information_complete == "" and best_query_webpage_information_similarity_score > 0.30:
+            webpage_information_complete = best_webpage_information
+        
+        state["information"] = webpage_information_complete[:3000]
+        state["best_query_webpage_information_similarity_score"] = best_query_webpage_information_similarity_score
+    except:
+        state["information"] = ""
+        state["best_query_webpage_information_similarity_score"] = -1.0
+    
+    # logger.info(f"Information: {state['information']}")
+    # logger.info(f"Information: {state['best_query_webpage_information_similarity_score']}")
 
     return state
 
@@ -859,6 +904,8 @@ class BasicAgent:
                 "messages": question,
             }
 
+            if len(tokenizer.encode(state["messages"][::-1])) < len(tokenizer.encode(state["messages"])):
+                state["messages"] = state["messages"][::-1]
 
             try:
                 response = self.safe_app.invoke(state)