Enhance tool descriptions and enforce stricter input/output format guidelines in `tools.py` and `agent.py`

src/gaia_solving_agent/agent.py

@@ -222,6 +222,8 @@ gaia_solving_agent = FunctionAgent(
     Try to get resources before querying them.
     If the analysis require a new external resource get it first.(e.g. a set of rules or a process)
     
+    When calling a tool, the inputs should be in a valid JSON format.
+    
     You will be provided a question, some known facts summarizing the user provided context and some sub-tasks to complete.
     You should follow the order of the sub-tasks.
     """,

src/gaia_solving_agent/tools.py

@@ -27,6 +27,7 @@ async def tavily_search_web(query: str) -> str:
     client = AsyncTavilyClient(api_key=TAVILY_API_KEY)
     return str(await client.search(query))
 
+
 async def vllm_ask_image_tool(ctx: Context, query: str) -> str:
     """
     Asynchronously processes a visual-linguistic query paired with image data
@@ -86,13 +87,35 @@ async def vllm_ask_image(query: str, images: ImageDocument | list[ImageDocument]
 simple_web_page_reader_tool = OnDemandLoaderTool.from_defaults(
     SimpleWebPageReader(html_to_text=True),
     name="simple_web_page_reader_tool",
-    description="Tool for loading content from a web page and return it as text",
+    description="""
+Tool for loading content from a web page and return it as text.
+Only the text content of the page is returned.
+
+You need to be provided with a URL.
+DO NOT GUESS ANY URL !
+If you don't have a URL in the user request, first use a browser tool to get one. 
+
+Do not use this tool for:
+- Web search
+- Wikipedia pages
+=> You have specialized tools for those needs.
+""",
 )
 simple_web_page_reader_toolspec = LoadAndSearchToolSpec.from_defaults(simple_web_page_reader_tool)
 
 youtube_transcript_reader_tool = OnDemandLoaderTool.from_defaults(
     YoutubeTranscriptReader(),
     name="youtube_transcript_reader_tool",
-    description="Tool for loading the audio transcript from a youtube video and return it as text",
+    description=r"""
+Tool for loading the audio transcript from a youtube video and return it as text.
+
+You must provide a youtube link in one of the following format:
+Supported formats include:
+- youtube.com/watch?v={video_id} (with or without 'www.')
+- youtube.com/embed?v={video_id} (with or without 'www.')
+- youtu.be{video_id} (never includes www subdomain)
+
+If you are provided with a youtube link in the wrong format, make it fit one the supported format.
+""",
 )
 youtube_transcript_reader_toolspec = LoadAndSearchToolSpec.from_defaults(youtube_transcript_reader_tool)