Upload 5 files

README.md

@@ -8,8 +8,23 @@ sdk_version: 5.25.2
 app_file: app.py
 pinned: false
 hf_oauth: true
-# optional, default duration is 8 hours/480 minutes. Max duration is 30 days/43200 minutes.
 hf_oauth_expiration_minutes: 480
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# GAIA Agent — Final Assignment
+
+A `smolagents` CodeAgent built to answer GAIA Level 1 questions for the HF Agents
+course final assignment. Uses HF Inference Providers as the LLM backend with
+Serper for web search, plus tools for browsing, Wikipedia, YouTube transcripts,
+GAIA file attachments, audio transcription, and image analysis.
+
+## Required Space secrets
+
+| Secret           | Purpose                                                  |
+| ---------------- | -------------------------------------------------------- |
+| `HF_TOKEN`       | HF Inference Providers (uses your Pro quota)             |
+| `SERPER_API_KEY` | Web search via serper.dev                                |
+
+Optional overrides:
+- `AGENT_MODEL_ID` — default `meta-llama/Llama-3.3-70B-Instruct`
+- `AGENT_PROVIDER` — default `auto` (HF Inference Providers routing)

agent.py

@@ -0,0 +1,112 @@
+"""GAIA agent: smolagents CodeAgent + HF Inference Providers backend."""
+from __future__ import annotations
+
+import os
+import re
+from typing import Optional
+
+from smolagents import CodeAgent, InferenceClientModel
+
+from tools import (
+    analyze_image,
+    download_task_file,
+    read_table,
+    read_webpage,
+    transcribe_audio,
+    web_search,
+    wikipedia_search,
+    youtube_transcript,
+)
+
+DEFAULT_MODEL_ID = "meta-llama/Llama-3.3-70B-Instruct"
+DEFAULT_PROVIDER = "auto"
+
+SYSTEM_PROMPT = """You are a careful research agent solving GAIA benchmark questions.
+
+Hard rules for the FINAL ANSWER:
+- Answers are graded by EXACT string match after light normalization. Be precise.
+- Output ONLY the answer — no prefixes, no units unless the question asks, no
+  trailing punctuation, no quotes.
+- Numbers: digits only unless the question wants words. No commas, no $, no %.
+- Strings: lowercase only if the question wants lowercase; otherwise match the
+  natural casing of the entity. Do not add articles (a, an, the).
+- Lists: comma + single space between items, in the order requested.
+
+Workflow:
+1. If the question references "the attached file", "this Excel/CSV/audio/image",
+   etc., call `download_task_file(task_id)` FIRST. The task_id is given to you.
+2. Pick the right tool: `read_table` for spreadsheets, `transcribe_audio` for
+   audio, `analyze_image` for images, `read_webpage` after `web_search` for the
+   open web, `wikipedia_search` for encyclopedic facts, `youtube_transcript`
+   for YouTube. Use Python freely for math, parsing, and logic.
+3. Cross-check facts when an answer feels uncertain. Don't fabricate.
+4. End with `final_answer("<the exact answer>")`.
+"""
+
+
+class GaiaAgent:
+    """Stateless wrapper around a smolagents CodeAgent."""
+
+    def __init__(self, model_id: Optional[str] = None, provider: Optional[str] = None):
+        token = os.getenv("HF_TOKEN")
+        if not token:
+            raise RuntimeError(
+                "HF_TOKEN is not set. Add it as a Space secret to use HF Inference."
+            )
+        self.model = InferenceClientModel(
+            model_id=model_id or os.getenv("AGENT_MODEL_ID", DEFAULT_MODEL_ID),
+            provider=provider or os.getenv("AGENT_PROVIDER", DEFAULT_PROVIDER),
+            token=token,
+            max_tokens=2048,
+        )
+        self.tools = [
+            web_search,
+            read_webpage,
+            wikipedia_search,
+            youtube_transcript,
+            download_task_file,
+            read_table,
+            transcribe_audio,
+            analyze_image,
+        ]
+
+    def _build_agent(self) -> CodeAgent:
+        # Fresh agent per question — keeps memory clean between tasks.
+        return CodeAgent(
+            tools=self.tools,
+            model=self.model,
+            max_steps=8,
+            additional_authorized_imports=[
+                "pandas", "numpy", "json", "re", "math", "statistics",
+                "itertools", "collections", "datetime", "csv", "io",
+                "pathlib", "string", "base64",
+            ],
+        )
+
+    def __call__(self, question: str, task_id: Optional[str] = None) -> str:
+        agent = self._build_agent()
+        prompt = SYSTEM_PROMPT
+        if task_id:
+            prompt += f"\nThe current task_id is: {task_id}\n"
+        prompt += f"\nQuestion:\n{question}\n"
+        try:
+            raw = agent.run(prompt)
+        except Exception as e:
+            return f"AGENT ERROR: {e}"
+        return self._normalize(str(raw))
+
+    @staticmethod
+    def _normalize(text: str) -> str:
+        """Clean common LLM cruft so the grader's exact match has a chance."""
+        if not text:
+            return text
+        t = text.strip()
+        # Strip "FINAL ANSWER:" prefix if the model left it in.
+        t = re.sub(r"^\s*final answer\s*[:\-]\s*", "", t, flags=re.IGNORECASE)
+        # Strip wrapping quotes / backticks.
+        if len(t) >= 2 and t[0] == t[-1] and t[0] in "\"'`":
+            t = t[1:-1].strip()
+        # Drop trailing sentence-ending punctuation unless the answer is a sentence.
+        if not re.search(r"\s", t) or t.count(" ") < 6:
+            t = t.rstrip(".")
+        return t.strip()

app.py

@@ -1,34 +1,22 @@
 import os
+import traceback
+
 import gradio as gr
-import requests
-import inspect
 import pandas as pd
+import requests
+
+from agent import GaiaAgent
 
-# (Keep Constants as is)
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
-# --- Basic Agent Definition ---
-# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
-class BasicAgent:
-    def __init__(self):
-        print("BasicAgent initialized.")
-    def __call__(self, question: str) -> str:
-        print(f"Agent received question (first 50 chars): {question[:50]}...")
-        fixed_answer = "This is a default answer."
-        print(f"Agent returning fixed answer: {fixed_answer}")
-        return fixed_answer
-
-def run_and_submit_all( profile: gr.OAuthProfile | None):
-    """
-    Fetches all questions, runs the BasicAgent on them, submits all answers,
-    and displays the results.
-    """
-    # --- Determine HF Space Runtime URL and Repo URL ---
-    space_id = os.getenv("SPACE_ID") # Get the SPACE_ID for sending link to the code
+
+def run_and_submit_all(profile: gr.OAuthProfile | None):
+    """Fetch GAIA questions, run the agent, submit answers, render results."""
+    space_id = os.getenv("SPACE_ID")
 
     if profile:
-        username= f"{profile.username}"
+        username = f"{profile.username}"
         print(f"User logged in: {username}")
     else:
         print("User not logged in.")
@@ -38,80 +26,93 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
     questions_url = f"{api_url}/questions"
     submit_url = f"{api_url}/submit"
 
-    # 1. Instantiate Agent ( modify this part to create your agent)
+    # 1. Instantiate agent
     try:
-        agent = BasicAgent()
+        agent = GaiaAgent()
     except Exception as e:
         print(f"Error instantiating agent: {e}")
         return f"Error initializing agent: {e}", None
-    # In the case of an app running as a hugging Face space, this link points toward your codebase ( usefull for others so please keep it public)
+
     agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
     print(agent_code)
 
-    # 2. Fetch Questions
+    # 2. Fetch questions
     print(f"Fetching questions from: {questions_url}")
     try:
-        response = requests.get(questions_url, timeout=15)
+        response = requests.get(questions_url, timeout=30)
         response.raise_for_status()
         questions_data = response.json()
         if not questions_data:
-             print("Fetched questions list is empty.")
-             return "Fetched questions list is empty or invalid format.", None
+            return "Fetched questions list is empty or invalid format.", None
         print(f"Fetched {len(questions_data)} questions.")
     except requests.exceptions.RequestException as e:
-        print(f"Error fetching questions: {e}")
         return f"Error fetching questions: {e}", None
     except requests.exceptions.JSONDecodeError as e:
-         print(f"Error decoding JSON response from questions endpoint: {e}")
-         print(f"Response text: {response.text[:500]}")
-         return f"Error decoding server response for questions: {e}", None
+        return f"Error decoding server response for questions: {e}", None
     except Exception as e:
-        print(f"An unexpected error occurred fetching questions: {e}")
         return f"An unexpected error occurred fetching questions: {e}", None
 
-    # 3. Run your Agent
+    # 3. Run agent over all questions
     results_log = []
     answers_payload = []
     print(f"Running agent on {len(questions_data)} questions...")
-    for item in questions_data:
+    for i, item in enumerate(questions_data, 1):
         task_id = item.get("task_id")
         question_text = item.get("question")
         if not task_id or question_text is None:
             print(f"Skipping item with missing task_id or question: {item}")
             continue
+        print(f"\n=== [{i}/{len(questions_data)}] task {task_id} ===")
         try:
-            submitted_answer = agent(question_text)
-            answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
-            results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
+            submitted_answer = agent(question_text, task_id=task_id)
         except Exception as e:
-             print(f"Error running agent on task {task_id}: {e}")
-             results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
+            traceback.print_exc()
+            submitted_answer = f"AGENT ERROR: {e}"
+        print(f"  -> {submitted_answer!r}")
+        answers_payload.append(
+            {"task_id": task_id, "submitted_answer": submitted_answer}
+        )
+        results_log.append(
+            {
+                "Task ID": task_id,
+                "Question": question_text,
+                "Submitted Answer": submitted_answer,
+            }
+        )
 
     if not answers_payload:
-        print("Agent did not produce any answers to submit.")
-        return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
+        return (
+            "Agent did not produce any answers to submit.",
+            pd.DataFrame(results_log),
+        )
 
-    # 4. Prepare Submission 
-    submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
-    status_update = f"Agent finished. Submitting {len(answers_payload)} answers for user '{username}'..."
-    print(status_update)
+    # 4. Prepare submission
+    submission_data = {
+        "username": username.strip(),
+        "agent_code": agent_code,
+        "answers": answers_payload,
+    }
+    print(
+        f"Agent finished. Submitting {len(answers_payload)} answers for "
+        f"user '{username}'..."
+    )
 
     # 5. Submit
     print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
     try:
-        response = requests.post(submit_url, json=submission_data, timeout=60)
+        response = requests.post(submit_url, json=submission_data, timeout=120)
         response.raise_for_status()
         result_data = response.json()
         final_status = (
             f"Submission Successful!\n"
             f"User: {result_data.get('username')}\n"
             f"Overall Score: {result_data.get('score', 'N/A')}% "
-            f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
+            f"({result_data.get('correct_count', '?')}/"
+            f"{result_data.get('total_attempted', '?')} correct)\n"
             f"Message: {result_data.get('message', 'No message received.')}"
         )
         print("Submission successful.")
-        results_df = pd.DataFrame(results_log)
-        return final_status, results_df
+        return final_status, pd.DataFrame(results_log)
     except requests.exceptions.HTTPError as e:
         error_detail = f"Server responded with status {e.response.status_code}."
         try:
@@ -119,42 +120,28 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
             error_detail += f" Detail: {error_json.get('detail', e.response.text)}"
         except requests.exceptions.JSONDecodeError:
             error_detail += f" Response: {e.response.text[:500]}"
-        status_message = f"Submission Failed: {error_detail}"
-        print(status_message)
-        results_df = pd.DataFrame(results_log)
-        return status_message, results_df
+        return f"Submission Failed: {error_detail}", pd.DataFrame(results_log)
     except requests.exceptions.Timeout:
-        status_message = "Submission Failed: The request timed out."
-        print(status_message)
-        results_df = pd.DataFrame(results_log)
-        return status_message, results_df
+        return "Submission Failed: The request timed out.", pd.DataFrame(results_log)
     except requests.exceptions.RequestException as e:
-        status_message = f"Submission Failed: Network error - {e}"
-        print(status_message)
-        results_df = pd.DataFrame(results_log)
-        return status_message, results_df
+        return f"Submission Failed: Network error - {e}", pd.DataFrame(results_log)
     except Exception as e:
-        status_message = f"An unexpected error occurred during submission: {e}"
-        print(status_message)
-        results_df = pd.DataFrame(results_log)
-        return status_message, results_df
+        return (
+            f"An unexpected error occurred during submission: {e}",
+            pd.DataFrame(results_log),
+        )
 
 
 # --- Build Gradio Interface using Blocks ---
 with gr.Blocks() as demo:
-    gr.Markdown("# Basic Agent Evaluation Runner")
+    gr.Markdown("# GAIA Agent — Final Assignment Runner")
     gr.Markdown(
         """
         **Instructions:**
-
-        1.  Please clone this space, then modify the code to define your agent's logic, the tools, the necessary packages, etc ...
-        2.  Log in to your Hugging Face account using the button below. This uses your HF username for submission.
-        3.  Click 'Run Evaluation & Submit All Answers' to fetch questions, run your agent, submit answers, and see the score.
-
-        ---
-        **Disclaimers:**
-        Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
-        This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
+        1. Add `HF_TOKEN` and `SERPER_API_KEY` as Space secrets.
+        2. Log in to Hugging Face with the button below.
+        3. Click **Run Evaluation & Submit All Answers**. Running 20 questions
+           can take 10–20 minutes; stay on the tab.
         """
     )
 
@@ -162,35 +149,29 @@ with gr.Blocks() as demo:
 
     run_button = gr.Button("Run Evaluation & Submit All Answers")
 
-    status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
-    # Removed max_rows=10 from DataFrame constructor
+    status_output = gr.Textbox(
+        label="Run Status / Submission Result", lines=5, interactive=False
+    )
     results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
 
-    run_button.click(
-        fn=run_and_submit_all,
-        outputs=[status_output, results_table]
-    )
+    run_button.click(fn=run_and_submit_all, outputs=[status_output, results_table])
+
 
 if __name__ == "__main__":
-    print("\n" + "-"*30 + " App Starting " + "-"*30)
-    # Check for SPACE_HOST and SPACE_ID at startup for information
+    print("\n" + "-" * 30 + " App Starting " + "-" * 30)
     space_host_startup = os.getenv("SPACE_HOST")
-    space_id_startup = os.getenv("SPACE_ID") # Get SPACE_ID at startup
+    space_id_startup = os.getenv("SPACE_ID")
 
     if space_host_startup:
-        print(f"✅ SPACE_HOST found: {space_host_startup}")
-        print(f"   Runtime URL should be: https://{space_host_startup}.hf.space")
-    else:
-        print("ℹ️  SPACE_HOST environment variable not found (running locally?).")
-
-    if space_id_startup: # Print repo URLs if SPACE_ID is found
-        print(f"✅ SPACE_ID found: {space_id_startup}")
-        print(f"   Repo URL: https://huggingface.co/spaces/{space_id_startup}")
-        print(f"   Repo Tree URL: https://huggingface.co/spaces/{space_id_startup}/tree/main")
+        print(f"SPACE_HOST: {space_host_startup}")
+        print(f"  Runtime URL: https://{space_host_startup}.hf.space")
     else:
-        print("ℹ️  SPACE_ID environment variable not found (running locally?). Repo URL cannot be determined.")
+        print("SPACE_HOST not set (running locally?).")
 
-    print("-"*(60 + len(" App Starting ")) + "\n")
+    if space_id_startup:
+        print(f"SPACE_ID: {space_id_startup}")
+        print(f"  Repo URL: https://huggingface.co/spaces/{space_id_startup}")
 
-    print("Launching Gradio Interface for Basic Agent Evaluation...")
-    demo.launch(debug=True, share=False)
+    print("-" * (60 + len(" App Starting ")) + "\n")
+    print("Launching Gradio Interface for GAIA Agent Runner...")
+    demo.launch(debug=True, share=False)

requirements.txt

@@ -1,2 +1,14 @@
-gradio
-requests
+gradio>=5.25.2
+requests
+pandas
+smolagents>=1.14.0
+huggingface_hub>=0.27.0
+duckduckgo-search
+wikipedia-api
+youtube-transcript-api
+beautifulsoup4
+markdownify
+openpyxl
+python-dotenv
+Pillow
+pypdf

tools.py

@@ -0,0 +1,450 @@
+"""Custom tools for the GAIA agent.
+
+Each tool is a @tool-decorated function that smolagents can call from a CodeAgent.
+Keep tool docstrings precise — the LLM reads them to decide when to call.
+"""
+from __future__ import annotations
+
+import io
+import os
+import re
+import tempfile
+from pathlib import Path
+from typing import Optional
+from urllib.parse import urlparse
+
+import requests
+from smolagents import tool
+
+DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
+USER_AGENT = (
+    "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
+    "(KHTML, like Gecko) Chrome/124.0 Safari/537.36"
+)
+
+
+# ---------------------------------------------------------------------------
+# Web search
+# ---------------------------------------------------------------------------
+@tool
+def web_search(query: str, num_results: int = 8) -> str:
+    """Search the web with Serper (Google results) and return the top hits.
+
+    Args:
+        query: The search query.
+        num_results: How many results to return (1-10).
+
+    Returns:
+        A text block of results: title, link, snippet. Use this to find URLs
+        worth reading with `read_webpage`.
+    """
+    api_key = os.getenv("SERPER_API_KEY")
+    num_results = max(1, min(int(num_results), 10))
+    if not api_key:
+        # Fallback to DuckDuckGo if no Serper key.
+        try:
+            from duckduckgo_search import DDGS
+
+            with DDGS() as ddgs:
+                hits = list(ddgs.text(query, max_results=num_results))
+            if not hits:
+                return "No results."
+            return "\n\n".join(
+                f"[{i + 1}] {h.get('title', '')}\n{h.get('href', '')}\n{h.get('body', '')}"
+                for i, h in enumerate(hits)
+            )
+        except Exception as e:  # pragma: no cover
+            return f"Search failed (no SERPER_API_KEY, DDG fallback errored): {e}"
+
+    try:
+        resp = requests.post(
+            "https://google.serper.dev/search",
+            headers={"X-API-KEY": api_key, "Content-Type": "application/json"},
+            json={"q": query, "num": num_results},
+            timeout=20,
+        )
+        resp.raise_for_status()
+        data = resp.json()
+    except Exception as e:
+        return f"Serper search failed: {e}"
+
+    parts: list[str] = []
+    if "answerBox" in data:
+        ab = data["answerBox"]
+        parts.append(
+            "ANSWER BOX:\n"
+            + (ab.get("answer") or ab.get("snippet") or ab.get("title") or "").strip()
+        )
+    if "knowledgeGraph" in data:
+        kg = data["knowledgeGraph"]
+        parts.append(
+            f"KNOWLEDGE GRAPH: {kg.get('title', '')} — {kg.get('description', '')}"
+        )
+    for i, item in enumerate(data.get("organic", [])[:num_results], 1):
+        parts.append(
+            f"[{i}] {item.get('title', '')}\n{item.get('link', '')}\n"
+            f"{item.get('snippet', '')}"
+        )
+    return "\n\n".join(parts) if parts else "No results."
+
+
+# ---------------------------------------------------------------------------
+# Web page reader
+# ---------------------------------------------------------------------------
+@tool
+def read_webpage(url: str, max_chars: int = 8000) -> str:
+    """Fetch a URL and return its main text content as Markdown.
+
+    Args:
+        url: The full URL to fetch (http or https).
+        max_chars: Maximum characters to return (truncated tail dropped).
+
+    Returns:
+        Markdown text. Use after `web_search` to actually read a page.
+    """
+    try:
+        from bs4 import BeautifulSoup
+        from markdownify import markdownify
+    except Exception as e:  # pragma: no cover
+        return f"Missing deps: {e}"
+
+    if not url.startswith(("http://", "https://")):
+        return f"Invalid URL: {url}"
+
+    try:
+        resp = requests.get(url, headers={"User-Agent": USER_AGENT}, timeout=25)
+        resp.raise_for_status()
+    except Exception as e:
+        return f"Fetch failed for {url}: {e}"
+
+    ctype = resp.headers.get("Content-Type", "").lower()
+    if "pdf" in ctype or url.lower().endswith(".pdf"):
+        return _pdf_to_text(resp.content, max_chars)
+
+    soup = BeautifulSoup(resp.text, "html.parser")
+    for tag in soup(["script", "style", "noscript", "header", "footer", "nav"]):
+        tag.decompose()
+    md = markdownify(str(soup), heading_style="ATX")
+    md = re.sub(r"\n{3,}", "\n\n", md).strip()
+    if len(md) > max_chars:
+        md = md[:max_chars] + "\n\n[...truncated...]"
+    return md
+
+
+def _pdf_to_text(data: bytes, max_chars: int) -> str:
+    try:
+        from pypdf import PdfReader
+    except Exception:
+        try:
+            from PyPDF2 import PdfReader  # type: ignore
+        except Exception as e:
+            return f"PDF read failed (install pypdf): {e}"
+    try:
+        reader = PdfReader(io.BytesIO(data))
+        text = "\n\n".join((p.extract_text() or "") for p in reader.pages)
+    except Exception as e:
+        return f"PDF parse failed: {e}"
+    if len(text) > max_chars:
+        text = text[:max_chars] + "\n\n[...truncated...]"
+    return text
+
+
+# ---------------------------------------------------------------------------
+# Wikipedia
+# ---------------------------------------------------------------------------
+@tool
+def wikipedia_search(query: str, sentences: int = 8) -> str:
+    """Look up a topic on English Wikipedia.
+
+    Args:
+        query: The page title or topic.
+        sentences: Sentences of summary to return.
+
+    Returns:
+        A summary block with the page URL, or an error message.
+    """
+    try:
+        import wikipediaapi
+    except Exception as e:  # pragma: no cover
+        return f"Missing deps: {e}"
+
+    wiki = wikipediaapi.Wikipedia(user_agent=USER_AGENT, language="en")
+    page = wiki.page(query)
+    if not page.exists():
+        # Try a search-then-fetch with the search API.
+        try:
+            resp = requests.get(
+                "https://en.wikipedia.org/w/api.php",
+                params={
+                    "action": "query",
+                    "list": "search",
+                    "srsearch": query,
+                    "format": "json",
+                    "srlimit": 1,
+                },
+                headers={"User-Agent": USER_AGENT},
+                timeout=15,
+            )
+            hits = resp.json().get("query", {}).get("search", [])
+            if not hits:
+                return f"No Wikipedia page found for: {query}"
+            page = wiki.page(hits[0]["title"])
+        except Exception as e:
+            return f"Wikipedia lookup failed: {e}"
+        if not page.exists():
+            return f"No Wikipedia page found for: {query}"
+
+    summary = page.summary
+    parts = re.split(r"(?<=[.!?])\s+", summary)
+    out = " ".join(parts[: max(1, int(sentences))])
+    return f"{page.title}\n{page.fullurl}\n\n{out}"
+
+
+# ---------------------------------------------------------------------------
+# YouTube transcript
+# ---------------------------------------------------------------------------
+@tool
+def youtube_transcript(url_or_id: str) -> str:
+    """Fetch the transcript of a YouTube video.
+
+    Args:
+        url_or_id: A full YouTube URL or just the 11-char video ID.
+
+    Returns:
+        Plain text transcript, or an error message.
+    """
+    vid = _yt_id(url_or_id)
+    if not vid:
+        return f"Could not parse YouTube id from: {url_or_id}"
+    try:
+        from youtube_transcript_api import YouTubeTranscriptApi
+    except Exception as e:  # pragma: no cover
+        return f"Missing deps: {e}"
+    try:
+        chunks = YouTubeTranscriptApi.get_transcript(vid)
+    except Exception as e:
+        return f"Transcript fetch failed: {e}"
+    return " ".join(c["text"] for c in chunks)
+
+
+def _yt_id(s: str) -> Optional[str]:
+    s = s.strip()
+    if re.fullmatch(r"[A-Za-z0-9_-]{11}", s):
+        return s
+    try:
+        u = urlparse(s)
+    except Exception:
+        return None
+    if u.hostname in ("youtu.be",):
+        return u.path.lstrip("/")[:11] or None
+    if u.hostname and "youtube" in u.hostname:
+        from urllib.parse import parse_qs
+
+        qs = parse_qs(u.query)
+        v = qs.get("v", [None])[0]
+        if v:
+            return v[:11]
+        m = re.search(r"/(embed|shorts)/([A-Za-z0-9_-]{11})", u.path)
+        if m:
+            return m.group(2)
+    m = re.search(r"([A-Za-z0-9_-]{11})", s)
+    return m.group(1) if m else None
+
+
+# ---------------------------------------------------------------------------
+# GAIA file attachment
+# ---------------------------------------------------------------------------
+@tool
+def download_task_file(task_id: str) -> str:
+    """Download the file attachment for a GAIA task (if one exists).
+
+    Args:
+        task_id: The task id of the current question.
+
+    Returns:
+        Absolute local path of the downloaded file, or a message saying
+        no file is attached. Read the file with normal Python after.
+    """
+    base = os.getenv("GAIA_API_URL", DEFAULT_API_URL).rstrip("/")
+    url = f"{base}/files/{task_id}"
+    try:
+        resp = requests.get(url, timeout=30)
+    except Exception as e:
+        return f"Download error: {e}"
+    if resp.status_code == 404:
+        return "NO_FILE: this task has no attachment."
+    if resp.status_code != 200:
+        return f"Download failed: HTTP {resp.status_code}"
+
+    name = _filename_from_response(resp, task_id)
+    out_dir = Path(tempfile.gettempdir()) / "gaia_files"
+    out_dir.mkdir(parents=True, exist_ok=True)
+    path = out_dir / name
+    path.write_bytes(resp.content)
+    return str(path.resolve())
+
+
+def _filename_from_response(resp: requests.Response, task_id: str) -> str:
+    cd = resp.headers.get("Content-Disposition", "")
+    m = re.search(r'filename\*?=(?:UTF-\d\'\')?"?([^";]+)"?', cd)
+    if m:
+        return m.group(1).strip()
+    ctype = resp.headers.get("Content-Type", "").split(";")[0].strip()
+    ext = {
+        "text/plain": ".txt",
+        "text/csv": ".csv",
+        "application/pdf": ".pdf",
+        "application/json": ".json",
+        "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet": ".xlsx",
+        "application/vnd.ms-excel": ".xls",
+        "application/x-python": ".py",
+        "image/png": ".png",
+        "image/jpeg": ".jpg",
+        "audio/mpeg": ".mp3",
+        "audio/wav": ".wav",
+        "audio/x-wav": ".wav",
+        "audio/mp4": ".m4a",
+        "video/mp4": ".mp4",
+    }.get(ctype, "")
+    return f"{task_id}{ext}"
+
+
+# ---------------------------------------------------------------------------
+# Excel / CSV reader (deterministic helper so the LLM doesn't have to handcraft)
+# ---------------------------------------------------------------------------
+@tool
+def read_table(file_path: str, sheet: Optional[str] = None, max_rows: int = 200) -> str:
+    """Read an Excel/CSV file and return a textual preview.
+
+    Args:
+        file_path: Absolute path to .xlsx / .xls / .csv / .tsv.
+        sheet: Optional sheet name (Excel only). Default: first sheet.
+        max_rows: Max rows to include in the preview.
+
+    Returns:
+        Column dtypes + a CSV-style preview. For deeper analysis, load it with
+        pandas yourself in a code block.
+    """
+    import pandas as pd
+
+    p = Path(file_path)
+    if not p.exists():
+        return f"File not found: {file_path}"
+    suffix = p.suffix.lower()
+    try:
+        if suffix in (".xlsx", ".xls"):
+            df = pd.read_excel(p, sheet_name=sheet or 0)
+        elif suffix == ".tsv":
+            df = pd.read_csv(p, sep="\t")
+        else:
+            df = pd.read_csv(p)
+    except Exception as e:
+        return f"Read failed: {e}"
+
+    head = df.head(max_rows)
+    info = [
+        f"shape: {df.shape}",
+        "dtypes:",
+        df.dtypes.astype(str).to_string(),
+        "",
+        "preview:",
+        head.to_csv(index=False),
+    ]
+    return "\n".join(info)
+
+
+# ---------------------------------------------------------------------------
+# Audio transcription via HF Inference (Whisper)
+# ---------------------------------------------------------------------------
+@tool
+def transcribe_audio(file_path: str) -> str:
+    """Transcribe an audio file (mp3/wav/m4a) using Whisper via HF Inference.
+
+    Args:
+        file_path: Absolute path to the audio file.
+
+    Returns:
+        The transcript text, or an error message.
+    """
+    from huggingface_hub import InferenceClient
+
+    token = os.getenv("HF_TOKEN")
+    if not token:
+        return "Missing HF_TOKEN for HF Inference."
+    p = Path(file_path)
+    if not p.exists():
+        return f"File not found: {file_path}"
+    model_id = os.getenv("ASR_MODEL_ID", "openai/whisper-large-v3")
+    try:
+        client = InferenceClient(token=token)
+        out = client.automatic_speech_recognition(p.read_bytes(), model=model_id)
+    except Exception as e:
+        return f"ASR failed: {e}"
+    if isinstance(out, dict):
+        return out.get("text", "")
+    return getattr(out, "text", str(out))
+
+
+# ---------------------------------------------------------------------------
+# Image VQA via HF Inference
+# ---------------------------------------------------------------------------
+@tool
+def analyze_image(file_path: str, question: str = "Describe this image in detail.") -> str:
+    """Ask a vision-language model about an image file.
+
+    Args:
+        file_path: Absolute path to a .png / .jpg / .jpeg / .webp file.
+        question: The question to ask about the image. Default: detailed description.
+
+    Returns:
+        The model's answer text.
+    """
+    import base64
+
+    from huggingface_hub import InferenceClient
+
+    token = os.getenv("HF_TOKEN")
+    if not token:
+        return "Missing HF_TOKEN for HF Inference."
+    p = Path(file_path)
+    if not p.exists():
+        return f"File not found: {file_path}"
+
+    model_id = os.getenv("VLM_MODEL_ID", "Qwen/Qwen2.5-VL-7B-Instruct")
+    provider = os.getenv("VLM_PROVIDER", "auto")
+
+    suffix = p.suffix.lower().lstrip(".")
+    mime = {"jpg": "jpeg"}.get(suffix, suffix) or "png"
+    b64 = base64.b64encode(p.read_bytes()).decode("ascii")
+    data_url = f"data:image/{mime};base64,{b64}"
+
+    try:
+        client = InferenceClient(token=token, provider=provider)
+        resp = client.chat.completions.create(
+            model=model_id,
+            messages=[
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": question},
+                        {"type": "image_url", "image_url": {"url": data_url}},
+                    ],
+                }
+            ],
+            max_tokens=512,
+        )
+        return resp.choices[0].message.content or ""
+    except Exception as e:
+        return f"VLM call failed: {e}"
+
+
+__all__ = [
+    "web_search",
+    "read_webpage",
+    "wikipedia_search",
+    "youtube_transcript",
+    "download_task_file",
+    "read_table",
+    "transcribe_audio",
+    "analyze_image",
+]