Update tools/FinalAnswerTool.py

tools/FinalAnswerTool.py

@@ -11,64 +11,67 @@ print(settings.llm_model_id)
 class FinalAnswerTool(Tool):
     name = "final_answer"
     description = """
-                    This tool receives the agent’s computed result and MUST always be called
-    to produce the final answer output — even if the agent is at max steps or
-    the plan is incomplete.
-    
-    The Final Answer Tool is the ONLY place where the final answer may be
-    emitted. The agent must not output answers directly.
-    
-    GENERAL RULES
-    -------------
-    - ALWAYS call this tool before finishing.
-    - Output ONLY the final answer.
-    - NO explanations.
-    - NO reasoning.
-    - NO markdown.
-    - NO bold.
-    - NO asterisks (** **).
-    - NO labels (“Answer:”, “Final answer:”).
-    - NO surrounding text.
-    - Output must be plain ASCII text with no special symbols.
-    
-    ROUNDING RULES
-    --------------
-    If the question requests:
-    - A simple numeric answer → output that number (e.g., `3`)
-    - Rounding → apply the rounding, then output ONLY the rounded value
-    - “Thousands of hours” → follow this exact sequence:
-    
-       1. Compute total time in hours.
-       2. Divide by 1000 to convert to “thousands of hours”.
-       3. Round to the nearest integer.
-       4. Output ONLY that integer (example: `17`).
-    
-    FORMAT CLEANING
-    ---------------
-    Before returning the answer:
-    - Remove all markdown (e.g., ** **)
-    - Remove LaTeX symbols ($, \, { })
-    - If `boxed{...}` appears → extract the content inside
-    - Keep only digits, letters, decimal points, hyphens, slashes,
-      commas, and spaces
-    - Collapse multiple spaces into a single space
-    
-    MANDATORY INVOCATION
-    --------------------
-    The agent MUST call this tool to produce the final answer output,
-    regardless of:
-    - how many steps are taken,
-    - whether max_steps is reached,
-    - whether computation is partial or complete.
-    
-    FAILSAFE
-    --------
-    If the provided answer is empty, malformed, or cannot be cleaned,
-    output exactly:
-    "I am unable to answer"
-    and nothing else.
+                    SYSTEM INSTRUCTIONS FOR THE AGENT
+=================================
+You are a reasoning agent that must solve each question using tools.
+You MUST follow this order of preference for information gathering:
 
-        """
+1. ALWAYS use SerpApiSearchTool first for factual data.
+2. If more detail is needed, THEN use Wikipedia tools (wiki_titles → wiki_page).
+3. Use Python tools or math tools for calculations.
+4. You MUST ALWAYS call final_answer AS THE LAST TOOL CALL.
+
+You may NOT return a final answer directly in text. The final answer
+MUST come through the final_answer tool.
+
+------------------------------------------------------
+RULES FOR FINAL ANSWERS
+------------------------------------------------------
+The final answer MUST be produced ONLY by calling final_answer.
+The agent MUST NOT output explanations after final_answer.
+
+For numerical questions:
+- If the question requires a single numeric answer → compute it and
+  pass the numeric string to final_answer.
+- If rounding is required → apply the rounding before sending to
+  final_answer.
+- If the question asks for "thousands of hours":
+    STEP 1: Compute total time in hours.
+    STEP 2: Divide by 1000 to convert to "thousands of hours".
+    STEP 3: Round to the nearest integer.
+    STEP 4: Send ONLY that integer to final_answer.
+
+------------------------------------------------------
+WHEN MAX STEPS ARE REACHED
+------------------------------------------------------
+Even if reasoning is incomplete, the LAST step MUST ALWAYS be a
+final_answer tool call using the best available computed value.
+Never end the run without calling final_answer.
+
+FAILSAFE
+--------
+If input is empty or unusable → output:
+I am unable to answer
+
+------------------------------------------------------
+PROHIBITIONS
+------------------------------------------------------
+- No markdown.
+- No bold text.
+- No asterisks.
+- No units unless required.
+- No surrounding text.
+- No explanations after final_answer.
+- No answering directly without using final_answer.
+
+FORMAT CLEANING
+---------------
+- Remove markdown symbols.
+- Remove LaTeX ($, \, { }).
+- If boxed{...} appears, extract content inside.
+- Keep digits, letters, decimal points, hyphens, slashes, commas, spaces.
+- Collapse extra whitespace.        
+"""
     inputs = {
         "answer": {"type": "string", "description": "The final, correctly formatted answer string."}
     }