tool improvement, smaller model

app.py

@@ -4,15 +4,12 @@ import requests
 import inspect
 import pandas as pd
 from smolagents import (
-    VisitWebpageTool, 
     DuckDuckGoSearchTool,
-    WikipediaSearchTool, 
     PythonInterpreterTool, 
     FinalAnswerTool, 
-    TransformersModel, 
     InferenceClientModel)
 from smolagents.agents import CodeAgent
-from tools import VisitWikiPageTool, SpeechToTextTool
+from tools import VisitWebpageTool, SpeechToTextTool
 from utils import QuestionLoader
 
 with open('system_prompt.txt', 'r') as file:
@@ -23,18 +20,26 @@ with open('system_prompt.txt', 'r') as file:
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 # Models
-MODEL = InferenceClientModel("Qwen/Qwen3-VL-30B-A3B-Instruct")
+#MODEL = InferenceClientModel("Qwen/Qwen3-VL-30B-A3B-Instruct")
+MODEL = InferenceClientModel("Qwen/Qwen3-VL-7B-Instruct")
 
 # --- Basic Agent Definition ---
 # ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
 class BasicAgent:
     def __init__(self):
-        print("BasicAgent initialized.")
-    def run(self, task: str, **kwargs) -> str:
-        print(f"Agent received question (first 50 chars): {task[:50]}...")
-        fixed_answer = "This is a default answer."
-        print(f"Agent returning fixed answer: {fixed_answer}")
-        return fixed_answer
+        self.agent = CodeAgent(
+            model=MODEL,
+            max_steps=10,
+            tools=[
+                DuckDuckGoSearchTool(),
+                VisitWebpageTool(),
+                SpeechToTextTool(),
+                PythonInterpreterTool(),
+                FinalAnswerTool()])
+        self.agent.prompt_templates["system_prompt"] = system_prompt
+
+    def run(self, **kwargs) -> str:
+        return self.agent.run(**kwargs)
     
 
 
@@ -61,18 +66,6 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
     # 1. Instantiate Agent ( modify this part to create your agent)
     try:
         agent = BasicAgent()
-        agent = CodeAgent(
-            model=MODEL,
-            max_steps=10,
-            tools=[
-                DuckDuckGoSearchTool(),
-                WikipediaSearchTool(),
-                VisitWebpageTool(),
-                VisitWikiPageTool(user_agent="hf-agent-course"),
-                SpeechToTextTool(),
-                PythonInterpreterTool(),
-                FinalAnswerTool()])
-        agent.prompt_templates["system_prompt"] = system_prompt
     except Exception as e:
         print(f"Error instantiating agent: {e}")
         return f"Error initializing agent: {e}", None

requirements.txt

@@ -13,3 +13,4 @@ pandas
 torch==2.9
 torchaudio
 torchcodec
+markdown-it-py==4.0.0

system_prompt.txt

@@ -1,4 +1,4 @@
-You are an expert assistant who can solve any task using code blobs. You will be given a task to solve as best you can.
+You are an expert assistant who can solve any task. You will be given a task to solve as best you can, you can also use code if you need to.
 To do so, you have been given access to a list of tools: these tools are basically Python functions which you can call with code.
 To solve the task, you must plan forward to proceed in a series of steps, in a cycle of Thought, Code, and Observation sequences.
 
@@ -172,13 +172,9 @@ specified otherwise.
 15. 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.
 16. Skip questions related to youtube videos since you do not have the tools to answer. Just answer 'Skip' in such cases.
+17. Wikipedia is your first to go website for facts and information retrieval.
+18. Don't over-use or over-rely on code to solve taks, use it when necessary.
 
-Here are suggestions, these are not rules, so you may decide to not follow them, however, they can make solving tasks easier:
-1. Some questions are related to a file, use the 'get_question_file' tool to retrieve the question's file content.
-2. The 'wikipedia_search' tool often returns incomplete results and is not able to parse tables. In order to get the full content of a wikipedia pages
-you should rely on the 'visit_wikipage' tool.
-3. To visit wikipedia pages you should use the 'visit_wikipage' tool and NOT the generic 'visit_webpage' tool since the latter won't work.
-4. Instead of using regex or code instruction to extract information from text it you are often better of relying on your own text understading capabilities.
 
 {%- if custom_instructions %}
 {{custom_instructions}}

tools.py

@@ -1,39 +1,51 @@
 from typing import Dict
 from transformers import pipeline
+from markdown_it import MarkdownIt
 from smolagents.tools import Tool
 import torchcodec
 
 
-class VisitWikiPageTool(Tool):
-    name = "visit_wikipage"
+class VisitWebpageTool(Tool):
+    name = "visit_webpage"
     description = (
-        "Visits a Wikipedia page at the given url and reads its content as a markdown string. Use this to browse Wikipedia wepages and get their full content."
+        "Visits a web page at the given url and reads its content as a markdown string and store it to a file"
     )
     inputs = {
         "url": {
             "type": "string",
             "description": "The url of the webpage to visit.",
         },
-        "max_length": {
-            "type": "integer",
-            "description": "Maximum number of characters to include in the response. Default 40000.",
-            "nullable": True
-        }
     }
     output_type = "string"
 
-    def __init__(self, user_agent: str):
+    def __init__(
+        self,
+        file_name: str = "web_content.md",
+        user_agent: str = "agent-course"
+        ):
         super().__init__()
+        self.file_name = file_name
         self.headers = {"User-Agent": user_agent}
 
-    def _truncate_content(self, content: str, max_length: int) -> str:
-        if len(content) <= max_length:
-            return content
-        return (
-            content[:max_length] + f"\n..._This content has been truncated to stay below {max_length} characters_...\n"
-        )
+    #def _truncate_content(self, content: str, max_length: int) -> str:
+    #    if len(content) <= max_length:
+    #        return content
+    #    return (
+    #        content[:max_length] + f"\n..._This content has been truncated to stay below {max_length} characters_...\n"
+    #    )
 
-    def forward(self, url: str, max_length: int = 40000) -> str:
+    def _inspect(self, doc: str) -> str:
+      mdit = MarkdownIt()
+      tokens = mdit.parse(doc)
+      content_table = ""
+      for token in tokens:
+        if token.type == "heading_open":
+          level = int(token.tag[-1]) - 1
+          text = token.map and tokens[tokens.index(token) + 1].content
+          content_table += "   " * level + text + "\n"
+      return content_table
+
+    def forward(self, url: str) -> str:
         try:
             import re
             import requests
@@ -50,11 +62,16 @@ class VisitWikiPageTool(Tool):
 
             # Convert the HTML content to Markdown
             markdown_content = markdownify(response.text).strip()
-            max_length = max_length if max_length is not None else 40000
+
             # Remove multiple line breaks
             markdown_content = re.sub(r"\n{3,}", "\n\n", markdown_content)
-            return self._truncate_content(markdown_content, max_length)
-
+            with open(self.file_name, "w") as f:
+              f.write(markdown_content)
+            try:
+              content_summary = self._inspect(markdown_content)
+              return f"Web page content saved in '{self.file_name}'. The content has the following section tree:\n {content_summary}. To read the full website content you can call 'read_mddoc('web_content.md')'"
+            except Exception:
+              return f"Web page content saved in {self.file_name}."
         except requests.exceptions.Timeout:
             return "The request timed out. Please try again later or check the URL."
         except RequestException as e:
@@ -62,6 +79,7 @@ class VisitWikiPageTool(Tool):
         except Exception as e:
             return f"An unexpected error occurred: {str(e)}"
         
+        
 class SpeechToTextTool(Tool):
   name = "transcriber"
   description = "This is a tool that transcribes an audio into text. It returns the transcribed text."
@@ -107,8 +125,91 @@ class SpeechToTextTool(Tool):
     self.pipe = pipeline("automatic-speech-recognition", model=model)
 
   def forward(self, audio_file: str, sample_rate: int=16000) -> str:
-    sample_rate = sample_rate if sample_rate is not None else 16000
-    with open(audio_file, "rb") as f:
-      decoder = torchcodec.decoders.AudioDecoder(f, sample_rate=sample_rate)
-      out = self.pipe(decoder)
-    return out["text"]
+    try:
+      sample_rate = sample_rate if sample_rate is not None else 16000
+      with open(audio_file, "rb") as f:
+        decoder = torchcodec.decoders.AudioDecoder(f, sample_rate=sample_rate)
+        audio_length = decoder.get_all_samples().data.shape[1]
+        out = self.pipe(decoder)
+      return out["text"]
+    except ValueError as e:
+        max_length = 300000
+        suggest_sample_rate = int(sample_rate * max_length/audio_length)
+        return f"The audio file to transcribe is too long, number of samples {audio_length}. You used a sample_rate of {sample_rate}, try using a smaller sample rate, like {suggest_sample_rate}"
+    except Exception as e:
+      raise e
+
+
+class ReadMdDoc(Tool):
+  name = "read_mddoc"
+  description = (
+        "Read an entire markdown file or a specific section of it."
+    )
+  inputs = {
+        "file_name": {
+            "type": "string",
+            "description": "The file to read it should have 'md' extension.",
+        },
+        "section": {
+            "type": "string",
+            "nullable": True,
+            "description": "If you want to read the entire file set this to 'all'. Otherwise you can look for a specific section title."
+        },
+        "max_length":{
+            "type": "integer",
+            "nullable": True,
+            "description": "The maximum number of characters to return if the content has more characters it will be truncated. Use 40000 as a default."
+        }
+    }
+  output_type = "string"
+
+  def __init__(self):
+    super().__init__()
+
+  def _truncate_content(self, content: str, max_length: int) -> str:
+      if len(content) <= max_length:
+          return content
+      return (
+          content[:max_length] + f"\n..._This content has been truncated to stay below {max_length} characters_...\n Does it have the information you need otherwise increase the max_length."
+      )
+
+  def get_token_map(self, tokens):
+    token_map = defaultdict(list)
+    stack = []
+    for i, token in enumerate(tokens):
+      if token.type == "heading_open":
+        text = token.map and tokens[tokens.index(token) + 1].content
+        token_map[text].append(i)
+        level = int(token.tag[-1])
+        while stack and level <= stack[-1][-1]:
+          key, _ = stack.pop()
+          token_map[key].append(i)
+        stack.append((text, level))
+    while stack:
+      text, _ = stack.pop()
+      token_map[text].append(i)
+    return token_map
+
+  def forward(
+      self,
+      file_name: str,
+      section: str = "all",
+      max_length: int = 40000):
+    try:
+      with open(file_name, "r") as f:
+        doc = f.read()
+    except FileNotFoundError:
+      return f"Can't find {file_name}, are you sure the file exists and that you have spelled it crrectly?"
+    try:
+      mdit = MarkdownIt()
+      tokens = mdit.parse(doc)
+    except Exception:
+      return "Error using the markdown parser, are you sure the file is in markdown format?"
+    token_map = self.get_token_map(tokens)
+    token_map["all"] = [0, len(tokens)]
+    if section in token_map:
+      start, end = tuple(token_map[section])
+      content = "\n".join([t.content for t in tokens[start:end]])
+      return self._truncate_content(content, max_length)
+    else:
+      return f"The required Section is not found in the document. The available sections are:\n {list(token_map.keys())}. If you don't see what you are looking for here, you can try returning all the document using setting argument section to 'all'"