Upload 6 files

README.md

@@ -8,8 +8,43 @@ 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 for the LLM and a
+small toolset for browsing, files, audio, and images. Self-consistency +
+verifier pass on top to maximize exact-match score.
+
+## Where things live
+
+| File          | Purpose                                                  |
+| ------------- | -------------------------------------------------------- |
+| `app.py`      | Gradio UI + run-and-submit pipeline                      |
+| `agent.py`    | `GaiaAgent` class — self-consistency + verifier          |
+| `tools.py`    | 8 tools (search, browse, wiki, YT, file, table, ASR, VLM)|
+| `config.py`   | **All non-secret settings** (model IDs, knobs)           |
+| `requirements.txt` | Python deps                                         |
+
+## Secrets (Space → Settings → Variables and secrets)
+
+| Secret           | Purpose                                                  |
+| ---------------- | -------------------------------------------------------- |
+| `HF_TOKEN`       | HF Inference Providers (uses your Pro quota)             |
+| `SERPER_API_KEY` | Web search via serper.dev                                |
+
+That's it. No model IDs in secrets — edit `config.py` for those.
+
+## Tuning
+
+Open `config.py` and change values directly:
+
+- `AGENT_MODEL_ID` — the LLM brain. Default DeepSeek-V3.
+- `SELF_CONSISTENCY_N` — independent attempts per question. 3 = balanced, 5 = max.
+- `MAX_STEPS` — agent reasoning steps per attempt.
+- `VLM_MODEL_ID` — vision model for image questions.
+
+Any of the constants in `config.py` can also be overridden by a Space env var
+of the same name, but that's optional.

agent.py

@@ -0,0 +1,267 @@
+"""GAIA agent: smolagents CodeAgent + HF Inference Providers backend.
+
+Pipeline per question:
+    1. Eager-download `/files/{task_id}` so attachments are never missed.
+    2. Run the CodeAgent N times (self-consistency).
+    3. Majority-vote across normalized candidate answers.
+    4. Verifier LLM pass: re-asks the model to pick the best answer in the
+       canonical GAIA `FINAL ANSWER:` format given the candidates.
+    5. Final string normalization for exact-match.
+"""
+from __future__ import annotations
+
+import os
+import re
+from collections import Counter
+from typing import Optional
+
+from huggingface_hub import InferenceClient
+from smolagents import CodeAgent, InferenceClientModel
+
+import config
+from tools import (
+    analyze_image,
+    download_task_file,
+    read_table,
+    read_webpage,
+    transcribe_audio,
+    web_search,
+    wikipedia_search,
+    youtube_transcript,
+)
+
+SYSTEM_PROMPT = """You are a general AI assistant solving questions from the GAIA benchmark.
+
+I will ask you a question. Think step by step using the tools available. When you
+are confident, finish your answer with the following template:
+
+FINAL ANSWER: [YOUR FINAL ANSWER]
+
+YOUR FINAL ANSWER must obey these rules exactly:
+- It is a number, OR as few words as possible, OR a comma separated list of
+  numbers and/or strings.
+- If a number: digits only. No commas inside numbers. No units ($, %, kg, etc.)
+  unless the question explicitly asks for them.
+- If a string: no articles (a, an, the), no abbreviations (write "Saint Louis"
+  not "St. Louis", "New York City" not "NYC"), and digits in plain text unless
+  the question asks for numerals.
+- If a comma separated list: items separated by ", " (comma + single space),
+  in the order requested, each item following the number/string rules above.
+- Do NOT include explanations, units, quotes, or trailing punctuation in the
+  FINAL ANSWER line itself.
+
+Tool playbook:
+- If the question mentions an attached file ("the attached", "this Excel /
+  CSV / audio / image / Python file", "the file"), call
+  `download_task_file(task_id)` FIRST to get a local path.
+- Spreadsheets / CSV: `read_table(path)` then pandas in Python.
+- Audio: `transcribe_audio(path)`.
+- Image: `analyze_image(path, question)` — phrase the question precisely.
+- YouTube link: `youtube_transcript(url)`.
+- Open web: `web_search(query)` then `read_webpage(url)` on the best hit.
+- Encyclopedic facts: `wikipedia_search(topic)`.
+- Use Python freely for math, parsing, sorting, set operations, etc.
+
+Self-check before answering:
+- Did you answer the literal question asked? (Not a related one.)
+- Does the format match the rules above?
+- If unsure, do one more verification search.
+"""
+
+VERIFIER_SYSTEM = """You are a strict GAIA answer formatter.
+
+You will see a GAIA question and several candidate answers produced by an agent.
+Pick the single best answer and reformat it to obey the GAIA answer format:
+
+- A number, a short string, or a comma-separated list.
+- Numbers: digits only, no commas, no units unless the question requires them.
+- Strings: no articles, no abbreviations, plain-text digits unless asked.
+- Comma-separated lists: ", " between items, in the order requested.
+- No explanations. No quotes. No trailing punctuation.
+
+Respond on a SINGLE LINE in exactly this template:
+FINAL ANSWER: <the answer>
+"""
+
+
+class GaiaAgent:
+    """CodeAgent + self-consistency + verifier."""
+
+    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._token = token
+        self._model_id = model_id or config.AGENT_MODEL_ID
+        self._provider = provider or config.AGENT_PROVIDER
+        self.model = InferenceClientModel(
+            model_id=self._model_id,
+            provider=self._provider,
+            token=token,
+            max_tokens=config.AGENT_MAX_TOKENS,
+        )
+        self.tools = [
+            web_search,
+            read_webpage,
+            wikipedia_search,
+            youtube_transcript,
+            download_task_file,
+            read_table,
+            transcribe_audio,
+            analyze_image,
+        ]
+        self._verifier = InferenceClient(token=token, provider=self._provider)
+        self._n = config.SELF_CONSISTENCY_N
+
+    # -------------------- public --------------------
+    def __call__(self, question: str, task_id: Optional[str] = None) -> str:
+        # 1) Pre-download attachment.
+        attachment_line = ""
+        if task_id:
+            try:
+                path = download_task_file(task_id)
+                if path and not path.startswith(
+                    ("NO_FILE", "Download error", "Download failed")
+                ):
+                    attachment_line = (
+                        f"An attached file for this task has already been "
+                        f"downloaded to: {path}\nUse the right tool on this "
+                        f"path (read_table for spreadsheets/CSV, "
+                        f"transcribe_audio for audio, analyze_image for "
+                        f"images, or open it with Python).\n"
+                    )
+            except Exception as e:
+                print(f"  pre-download skipped: {e}")
+
+        # 2) Self-consistency: N independent CodeAgent runs.
+        candidates: list[str] = []
+        for i in range(self._n):
+            try:
+                raw = self._single_run(question, task_id, attachment_line)
+            except Exception as e:
+                print(f"  attempt {i + 1} errored: {e}")
+                continue
+            norm = self._normalize(str(raw))
+            print(f"  attempt {i + 1}: {norm!r}")
+            if norm and not norm.startswith("AGENT ERROR"):
+                candidates.append(norm)
+
+        if not candidates:
+            return "AGENT ERROR: no successful attempts"
+
+        # 3) Majority vote on normalized strings (case-insensitive bucket).
+        voted = self._vote(candidates)
+
+        # 4) Verifier pass — pick + reformat using the LLM.
+        try:
+            verified = self._verify(question, candidates, voted)
+        except Exception as e:
+            print(f"  verifier errored, falling back to vote: {e}")
+            verified = voted
+
+        return self._normalize(verified)
+
+    # -------------------- helpers --------------------
+    def _build_agent(self) -> CodeAgent:
+        kwargs = dict(
+            tools=self.tools,
+            model=self.model,
+            max_steps=config.MAX_STEPS,
+            additional_authorized_imports=[
+                "pandas", "numpy", "json", "re", "math", "statistics",
+                "itertools", "collections", "datetime", "csv", "io",
+                "pathlib", "string", "base64", "urllib", "unicodedata",
+            ],
+        )
+        if config.PLANNING_INTERVAL > 0:
+            try:
+                return CodeAgent(planning_interval=config.PLANNING_INTERVAL, **kwargs)
+            except TypeError:
+                pass
+        return CodeAgent(**kwargs)
+
+    def _single_run(
+        self,
+        question: str,
+        task_id: Optional[str],
+        attachment_line: str,
+    ) -> str:
+        agent = self._build_agent()
+        prompt = SYSTEM_PROMPT
+        if task_id:
+            prompt += f"\nThe current task_id is: {task_id}\n"
+        if attachment_line:
+            prompt += attachment_line
+        prompt += f"\nQuestion:\n{question}\n"
+        return agent.run(prompt)
+
+    @staticmethod
+    def _vote(candidates: list[str]) -> str:
+        """Majority-vote with case-insensitive bucketing; ties → first seen."""
+        buckets: dict[str, list[str]] = {}
+        for c in candidates:
+            key = c.lower().strip()
+            buckets.setdefault(key, []).append(c)
+        ranked = sorted(
+            buckets.items(),
+            key=lambda kv: (-len(kv[1]), candidates.index(kv[1][0])),
+        )
+        return ranked[0][1][0]
+
+    def _verify(
+        self,
+        question: str,
+        candidates: list[str],
+        voted: str,
+    ) -> str:
+        """Ask the LLM to choose + reformat the best candidate."""
+        cand_block = "\n".join(f"- {c}" for c in candidates)
+        user = (
+            f"Question:\n{question}\n\n"
+            f"Candidate answers from independent attempts:\n{cand_block}\n\n"
+            f"Most common candidate: {voted}\n\n"
+            f"Return the best final answer in the canonical GAIA format."
+        )
+        resp = self._verifier.chat.completions.create(
+            model=self._model_id,
+            messages=[
+                {"role": "system", "content": VERIFIER_SYSTEM},
+                {"role": "user", "content": user},
+            ],
+            max_tokens=256,
+            temperature=0.0,
+        )
+        text = (resp.choices[0].message.content or "").strip()
+        m = re.search(r"final answer\s*[:\-]\s*(.+)", text, flags=re.IGNORECASE)
+        return m.group(1).strip() if m else text
+
+    @staticmethod
+    def _normalize(text: str) -> str:
+        """Strip every common LLM cruft pattern so exact-match has a chance."""
+        if not text:
+            return text
+        t = str(text).strip()
+
+        # 1) If the model emitted "FINAL ANSWER: X" anywhere, keep only X.
+        m = re.search(r"final answer\s*[:\-]\s*(.+)", t, flags=re.IGNORECASE | re.DOTALL)
+        if m:
+            t = m.group(1).strip()
+            t = t.split("\n")[0].strip()
+
+        # 2) Strip wrapping markdown bold/italic/code.
+        t = re.sub(r"^[*_`]+|[*_`]+$", "", t).strip()
+
+        # 3) Strip wrapping quotes (incl. smart quotes).
+        if len(t) >= 2 and t[0] == t[-1] and t[0] in "\"'`“”‘’":
+            t = t[1:-1].strip()
+
+        # 4) Drop a single trailing period if the answer is short (not a sentence).
+        if t.count(" ") < 6:
+            t = t.rstrip(".")
+
+        # 5) Collapse internal whitespace.
+        t = re.sub(r"[ \t]+", " ", t).strip()
+
+        return t

app.py

@@ -1,117 +1,116 @@
 import os
+import traceback
+
 import gradio as gr
-import requests
-import inspect
 import pandas as pd
+import requests
+
+import config
+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.")
         return "Please Login to Hugging Face with the button.", None
 
-    api_url = DEFAULT_API_URL
+    api_url = config.GAIA_API_URL
     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 +118,28 @@ 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 +147,29 @@ 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)

config.py

@@ -0,0 +1,66 @@
+"""Non-secret configuration for the GAIA agent.
+
+WHAT LIVES WHERE
+- Secrets (Space → Settings → Variables and secrets):
+    * HF_TOKEN          — required, for HF Inference Providers
+    * SERPER_API_KEY    — required, for web search
+- Everything else (model IDs, providers, step limits, tool knobs) lives in
+  this file. Edit values here and push — no secret changes needed.
+
+OVERRIDE RULES
+Any env var with the same name as a constant below will override the value at
+import time. This lets you A/B test from Space secrets without editing code.
+"""
+from __future__ import annotations
+
+import os
+
+
+def _env(name: str, default: str) -> str:
+    return os.getenv(name, default)
+
+
+# ---------------------------------------------------------------------------
+# LLM that powers the CodeAgent (the "brain")
+# ---------------------------------------------------------------------------
+# Strong open-source options ranked by GAIA performance (as of 2026):
+#   - "deepseek-ai/DeepSeek-V3-0324"        (provider="together")  ← default
+#   - "Qwen/Qwen2.5-72B-Instruct"           (provider="auto")
+#   - "meta-llama/Llama-3.3-70B-Instruct"   (provider="auto")
+#   - "deepseek-ai/DeepSeek-R1-Distill-Llama-70B" (reasoning-tuned)
+AGENT_MODEL_ID = _env("AGENT_MODEL_ID", "deepseek-ai/DeepSeek-V3-0324")
+AGENT_PROVIDER = _env("AGENT_PROVIDER", "together")
+AGENT_MAX_TOKENS = int(_env("AGENT_MAX_TOKENS", "4096"))
+
+# ---------------------------------------------------------------------------
+# Vision-language model (used by `analyze_image`)
+# ---------------------------------------------------------------------------
+VLM_MODEL_ID = _env("VLM_MODEL_ID", "Qwen/Qwen2.5-VL-72B-Instruct")
+VLM_PROVIDER = _env("VLM_PROVIDER", "auto")
+
+# ---------------------------------------------------------------------------
+# Speech-to-text model (used by `transcribe_audio`)
+# ---------------------------------------------------------------------------
+ASR_MODEL_ID = _env("ASR_MODEL_ID", "openai/whisper-large-v3")
+
+# ---------------------------------------------------------------------------
+# Agent loop knobs
+# ---------------------------------------------------------------------------
+# Max reasoning steps before the agent must commit to a final answer.
+MAX_STEPS = int(_env("AGENT_MAX_STEPS", "15"))
+# How often to interleave a planning step (0 = disabled).
+PLANNING_INTERVAL = int(_env("AGENT_PLANNING_INTERVAL", "4"))
+# Self-consistency: independent attempts per question. 3 = balanced, 5 = max.
+SELF_CONSISTENCY_N = int(_env("GAIA_SELF_CONSISTENCY_N", "3"))
+
+# ---------------------------------------------------------------------------
+# Tool defaults
+# ---------------------------------------------------------------------------
+SEARCH_RESULTS = int(_env("WEB_SEARCH_RESULTS", "10"))
+PAGE_MAX_CHARS = int(_env("READ_WEBPAGE_MAX_CHARS", "15000"))
+WIKI_SENTENCES = int(_env("WIKI_SENTENCES", "8"))
+
+# ---------------------------------------------------------------------------
+# GAIA scoring API
+# ---------------------------------------------------------------------------
+GAIA_API_URL = _env("GAIA_API_URL", "https://agents-course-unit4-scoring.hf.space")

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 +1,451 @@
+"""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
+
+import config
+
+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 = config.SEARCH_RESULTS) -> 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 = config.PAGE_MAX_CHARS) -> 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 = config.WIKI_SENTENCES) -> 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 = config.GAIA_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 = config.ASR_MODEL_ID
+    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 = config.VLM_MODEL_ID
+    provider = config.VLM_PROVIDER
+
+    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",
+]