improve prompt

agents/__init__.py

@@ -0,0 +1,45 @@
+from smolagents import (
+    DuckDuckGoSearchTool,
+    VisitWebpageTool,
+    WikipediaSearchTool,
+)
+from tools.text_search import TextSearch
+from tools.text_splitter import text_splitter
+from tools.webpage_parser import WebpageParser
+from tools.parse_wikipedia_table import WikipediaParser
+from tools.open_files import OpenFilesTool
+
+DEFAULT_ARGS = myagent_args = {
+    "provider": "litellm",
+    "model_id": "gemini/gemini-2.0-flash-lite",
+    # "api_base": OLLAMA_API_BASE,
+    "planning_interval": 3,
+    "add_base_tools": True,
+    "tools": [
+        DuckDuckGoSearchTool(),
+        WikipediaParser(),
+        VisitWebpageTool(),
+        TextSearch(),
+        text_splitter,
+        WikipediaSearchTool(
+            content_type="text",
+            extract_format="HTML"
+        ),
+        WebpageParser(),
+        OpenFilesTool(),
+    ],
+    "additional_authorized_imports": [
+        "pandas",
+        "numpy",
+        "datetime",
+        "json",
+        "re",
+        "math",
+        "os",
+        "requests",
+        "csv",
+        "urllib",
+    ],
+    "num_ctx": 128_000,
+    "temperature": 0.2,
+}

agents/agent.py

@@ -86,6 +86,6 @@ class MyAgent:
         """
 
         final_answer = self.agent.run(question)
-        print(f"Agent received question (first 50 chars): {question[:50]}...")
+        print(f"Agent received question (last 50 chars): {question[-50:]}...")
         print(f"Agent returning fixed answer: {final_answer}")
         return final_answer

app.py

@@ -6,49 +6,11 @@ from agents.agent import MyAgent
 import time
 from tqdm import tqdm
 from prompts.default_prompt import generate_prompt
-from smolagents import (
-    DuckDuckGoSearchTool,
-    VisitWebpageTool,
-)
-from tools.text_search import TextSearch
-from tools.text_splitter import text_splitter
-from tools.webpage_parser import WebpageParser
-from tools.parse_wikipedia_table import WikipediaParser
-from tools.open_files import OpenFilesTool
+from agents import DEFAULT_ARGS
 
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 
-myagent_args = {
-    "provider": "litellm",
-    "model_id": "gemini/gemini-2.0-flash-lite",
-    # "api_base": OLLAMA_API_BASE,
-    "planning_interval": 3,
-    "tools": [
-        DuckDuckGoSearchTool(),
-        WikipediaParser(),
-        VisitWebpageTool(),
-        TextSearch(),
-        text_splitter,
-        WebpageParser(),
-        OpenFilesTool(),
-    ],
-    "additional_authorized_imports": [
-        "pandas",
-        "numpy",
-        "datetime",
-        "json",
-        "re",
-        "math",
-        "os",
-        "requests",
-        "csv",
-        "urllib",
-    ],
-    "num_ctx": 8192,
-    "temperature": 0.2,
-}
-
 
 def run_and_submit_all(profile: gr.OAuthProfile | None):
     """
@@ -71,7 +33,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
 
     # 1. Instantiate Agent ( modify this part to create your agent)
     try:
-        agent = MyAgent(**myagent_args)
+        agent = MyAgent(**DEFAULT_ARGS)
 
     except Exception as e:
         print(f"Error instantiating agent: {e}")

prompts/default_prompt.py

@@ -15,20 +15,14 @@ def generate_prompt(question_text, file_name):
     When given a question:
     - If necessary, perform a web search using the tool `DuckDuckGoSearchTool` to find possible sources of information.
     - Use the `visit_webpage` tool to visit the webpage and extract the content in markdown format.
-    - If the web search only returns titles and short snippets, you MUST visit the actual webpage to read the full content before answering.
-    - Use the `WikipediaParser` tool to fetch and read the Wikipedia page when necessary.
-    - You just have the ability to read Wikipedia pages only.
-    - If the task requires reading, listening, or analyzing a file, you must use the file specified in the `file_name` field of the task metadata, not the file name mentioned casually inside the question text.
-    - Comma separated lists MUST contain a single space after each comma.
-    - If you are asked for a number, don't use comma to write your number neither use units such as $ or percent sign unless specified otherwise.
-    - If you are asked for a string, don't use articles, neither abbreviations (e.g. for cities), and write the digits in plain text unless specified otherwise.
-    - If you are asked for a comma separated list, apply the above rules depending of whether the element to be put in the list is a number or a string.
-    - Only answer after you have gathered enough information by reading the actual page contents.
+    - Use the `WikipediaSearchTool` to search for any information on Wikipedia, this will return HTML content. You need to then use the `WikipediaParser` tool to parse the HTML content into a clean, readable text format.
+    - If the file_name provided ends in ".py", use the `PythonInterpreterTool` to execute the code in the file and return the output.
+    - Use the `PythonInterpreterTool` to execute any Python code snippets you generate.
+    - Use the `TextSearch` tool to search for a substring within a string.
+    - Use the `text_splitter` tool to split a string into smaller chunks of text.
+    - If the task requires reading, listening, or analyzing a file, you must use the file specified in the `file_name` field of the task metadata, not the file name mentioned casually inside the question text. Use the `OpenFilesTool` to open the file and read its content.
     - Once you have the final answer, you must call `final_answer("your_answer")` immediately after printing it.
     - Do not retry or execute anything else after calling `final_answer`.
-    - `final_answer` must wrap the exact printed value.
-    Provide ONLY the precise answer requested. 
-    Do not include explanations, steps, reasoning, or additional text. 
     Be direct and specific. GAIA benchmark requires exact matching answers.
     Example: if asked "What is the capital of France?", respond exactly:
     Thoughts: I need to retrieve the capital of France from Wikipedia and output it directly.

run_local_agent.py

@@ -11,6 +11,7 @@ from tools.webpage_parser import WebpageParser
 from tools.parse_wikipedia_table import WikipediaParser
 from tools.open_files import OpenFilesTool
 from prompts.default_prompt import generate_prompt
+from agents import DEFAULT_ARGS
 
 
 import os
@@ -56,10 +57,10 @@ myagent_args = {
     "temperature": 0.2,
 }
 
-print(f"Using args: {myagent_args}")
+print(f"Using args: {DEFAULT_ARGS}")
 
 if __name__ == "__main__":
-    agent = MyAgent(**myagent_args)
+    agent = MyAgent(**DEFAULT_ARGS)
 
     with open(QUESTIONS_FILEPATH, "r") as f:
         questions = json.load(f)

tools/parse_wikipedia_table.py

@@ -26,7 +26,7 @@ class WikipediaParser(Tool):
         """
 
         headers = {
-            "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)"  # AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"
+            "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"
         }
         resp = requests.get(url, headers=headers, timeout=30)
         resp.raise_for_status()
@@ -49,7 +49,7 @@ class WikipediaParser(Tool):
             elif elem.name == "table":
                 elements.append(self.parse_wikipedia_table(elem))
 
-        return "\n".join(elements)
+        return "\n\n".join(elements)
 
     def parse_wikipedia_table(table: Tag) -> str:
         """