Update agent.py

agent.py

@@ -59,22 +59,25 @@ class GaiaAgent:
         Your task is to carefully and accurately answer questions by using the search tool when necessary.
         Always provide a complete and correct answer based on the information you find.
 
+        You must follow a Thought, Tool, Observation, Answer (TTOA) pattern.
+        
+        **Thought:** First, carefully consider the task. What information do you need to answer the question? Do you need to use a tool?
+        **Tool:** If you need to search, use the search_tavily tool. The format is: <TOOL_CODE>search_tavily("your search query")</TOOL_CODE>
+        **Observation:** After a tool call, you will receive an observation (the tool's output).
+        **Answer:** Once you have gathered all necessary information, provide your final, concise answer directly.
+
         Your available tools:
         1. search_tavily(query: str): Searches on Tavily and returns relevant results.
-           Use this tool to find information on the internet that you don't know or need to verify.
-
-        To use a tool, write it in the following exact format:
-        <TOOL_CODE>
-        tool_name("your search query")
-        </TOOL_CODE>
-        Example:
-        If you need to know the capital of France:
-        <TOOL_CODE>
-        search_tavily("capital of France")
-        </TOOL_CODE>
-
-        When you have found all the necessary information and are ready to answer the task, provide your final answer.
-        
+
+        Example Interaction:
+        Task: What is the capital of France?
+        Thought: I need to find the capital of France. I should use the search_tavily tool.
+        Tool: <TOOL_CODE>search_tavily("capital of France")</TOOL_CODE>
+        Observation: The capital of France is Paris.
+        Answer: The capital of France is Paris.
+
+        Now, let's start.
+
         Task: {task_description}
         """
 
@@ -82,25 +85,31 @@ class GaiaAgent:
         current_response = ""
 
         for i in range(max_iterations):
-            full_prompt = prompt + current_response + "\n\nWhat is the next step or your final answer?"
+            # Lägg till "Thought:" här för att uppmuntra modellen att starta sin tankeprocess
+            full_prompt = prompt + current_response + "\n\nThought:" 
             
             print(f"[{i+1}/{max_iterations}] Generating response with prompt length: {len(full_prompt)}")
             
             generated_text = self.text_generator(
                 full_prompt, 
-                max_new_tokens=1024, # Behold 1024 eller öka om behövs
+                max_new_tokens=1024, # Fortsätt med 1024 eller öka till 2048
                 num_return_sequences=1,
                 pad_token_id=self.tokenizer.eos_token_id,
                 do_sample=True,
                 top_k=50, top_p=0.95,
-                temperature=0.8 # Behold 0.8 eller justera vid behov
+                temperature=0.7 # Justera temperaturen till 0.7
             )[0]['generated_text']
 
             new_content = generated_text[len(full_prompt):].strip()
             print(f"DEBUG - Full generated_text: \n---START---\n{generated_text}\n---END---")
             print(f"DEBUG - Extracted new_content: '{new_content}'")
 
-            if "<TOOL_CODE>" in new_content and "</TOOL_CODE>" in new_content:
+            # Kontrollera om modellen genererade ett svar som en 'Answer:'
+            if "Answer:" in new_content:
+                final_answer = new_content.split("Answer:", 1)[1].strip()
+                print(f"Final answer from model:\n{final_answer}")
+                return final_answer
+            elif "<TOOL_CODE>" in new_content and "</TOOL_CODE>" in new_content:
                 start_index = new_content.find("<TOOL_CODE>") + len("<TOOL_CODE>")
                 end_index = new_content.find("</TOOL_CODE>")
                 tool_call_str = new_content[start_index:end_index].strip()
@@ -112,18 +121,18 @@ class GaiaAgent:
                         query = tool_call_str[len("search_tavily("):-1].strip().strip('"').strip("'")
                         tool_output = search_tavily(query)
                         print(f"Tool result: {tool_output[:200]}...")
-                        current_response += f"\n\nTool Result from {tool_call_str}:\n{tool_output}\n"
+                        current_response += f"\n\nObservation: {tool_output}\n"
                     else:
                         tool_output = f"Unknown tool: {tool_call_str}"
                         print(f"Error: {tool_output}")
-                        current_response += f"\n\n{tool_output}\n"
+                        current_response += f"\n\nObservation: {tool_output}\n"
                 except Exception as tool_e:
                     tool_output = f"Error running tool {tool_call_str}: {tool_e}"
                     print(f"Error: {tool_output}")
-                    current_response += f"\n\n{tool_output}\n"
+                    current_response += f"\n\nObservation: {tool_output}\n"
             else:
-                final_answer = new_content
-                print(f"Final answer from model:\n{final_answer}")
-                return final_answer.strip()
+                # Om modellen varken ger svar eller verktygskall men genererar något annat
+                current_response += f"\n\n{new_content}\n"
+                print(f"Model generated non-tool/non-answer content. Appending: {new_content[:100]}...")
 
-        return "Agent could not complete the task within the allowed iterations. Latest response: " + new_content.strip()
+        return "Agent could not complete the task within the allowed iterations. Latest response: " + new_content.strip() if new_content else "Agent could not complete the task within the allowed iterations. No meaningful content generated."