""" Agent GAIA – Unité 4 du Hugging Face AI Agents Course Sarra Laouini Architecture LangGraph : START → load_and_extract → assistant ⇄ tools → format_answer → END - load_and_extract : télécharge le fichier joint et le transforme en texte (Excel → pandas, MP3 → Whisper, image → VLM, code → texte) - assistant : LLM Groq (gpt-oss) avec outils, budget de 6 tours d'outils - tools : web_search (Tavily), wiki_rag (RAG LlamaIndex + bge-small, découpage par sections), python, youtube_transcript - format_answer : extrait la réponse après "FINAL ANSWER:" pour l'exact match Secrets nécessaires : GROQ_API_KEY, TAVILY_API_KEY, HF_TOKEN Option : MODEL_ID (défaut openai/gpt-oss-120b) """ import os import re import sys import json import uuid import base64 import mimetypes import subprocess from typing import TypedDict, Annotated import requests import pandas as pd from groq import Groq from tavily import TavilyClient from markdownify import markdownify from langchain_core.tools import tool from langchain_core.messages import (AnyMessage, AIMessage, HumanMessage, SystemMessage, ToolMessage) from langchain_groq import ChatGroq from langgraph.graph import StateGraph, START, END from langgraph.graph.message import add_messages from langgraph.prebuilt import ToolNode, tools_condition from llama_index.core import Document, VectorStoreIndex, Settings from llama_index.core.node_parser import SentenceSplitter from llama_index.embeddings.huggingface import HuggingFaceEmbedding API_URL = "https://agents-course-unit4-scoring.hf.space" FILES_DIR = "/tmp/gaia_files" MODEL_ID = os.getenv("MODEL_ID", "openai/gpt-oss-120b") VISION_MODEL = "qwen/qwen3.8-27b" HEADERS = {"User-Agent": "gaia-course-agent/1.0"} client = Groq() # lit GROQ_API_KEY # ============================================================================= # 1. State # ============================================================================= class AgentState(TypedDict): task_id: str question: str file_name: str file_path: str file_content: str messages: Annotated[list[AnyMessage], add_messages] answer: str # ============================================================================= # 2. Fichiers joints (nœud load_and_extract) # ============================================================================= _GAIA_FILES = None def gaia_files() -> dict: """task_id → chemin du fichier dans le dataset GAIA (si le serveur du cours renvoie 404).""" global _GAIA_FILES if _GAIA_FILES is None: from datasets import load_dataset from huggingface_hub import snapshot_download data_dir = snapshot_download(repo_id="gaia-benchmark/GAIA", repo_type="dataset", token=os.getenv("HF_TOKEN")) val = load_dataset(data_dir, "2023_level1", split="validation") # On ne lit QUE le chemin du fichier, jamais la réponse. _GAIA_FILES = {ex["task_id"]: os.path.join(data_dir, ex["file_path"]) for ex in val if ex["file_path"]} return _GAIA_FILES def download_file(task_id, file_name): os.makedirs(FILES_DIR, exist_ok=True) path = os.path.join(FILES_DIR, file_name) if os.path.exists(path): return path r = requests.get(f"{API_URL}/files/{task_id}", timeout=60) if r.status_code == 200: with open(path, "wb") as f: f.write(r.content) return path files = gaia_files() if task_id in files: print(f" ↳ serveur {r.status_code}, fichier pris dans le dataset GAIA") return files[task_id] raise FileNotFoundError(f"{file_name} introuvable (serveur {r.status_code})") def transcribe(path): with open(path, "rb") as f: out = client.audio.transcriptions.create(file=(os.path.basename(path), f.read()), model="whisper-large-v3") return out.text def describe_image(path, question): mime = mimetypes.guess_type(path)[0] or "image/png" with open(path, "rb") as f: b64 = base64.b64encode(f.read()).decode() params = dict( model=VISION_MODEL, temperature=0, max_tokens=800, messages=[{"role": "user", "content": [ {"type": "text", "text": f"Describe precisely and concisely everything in this image that helps answer: {question}"}, {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}}, ]}], ) try: out = client.chat.completions.create(**params, reasoning_effort="none") except Exception: out = client.chat.completions.create(**params) text = out.choices[0].message.content or "" return re.sub(r".*?", "", text, flags=re.DOTALL).strip() def extract(path, question): ext = os.path.splitext(path)[1].lower() if ext in (".py", ".txt", ".md", ".json", ".csv"): with open(path, encoding="utf-8", errors="ignore") as f: return f.read() if ext in (".xlsx", ".xls"): sheets = pd.read_excel(path, sheet_name=None) return "\n\n".join(f"Sheet '{n}':\n{df.to_csv(index=False)}" for n, df in sheets.items()) if ext == ".pdf": from pypdf import PdfReader return "\n".join(p.extract_text() or "" for p in PdfReader(path).pages) if ext in (".mp3", ".wav", ".m4a"): return transcribe(path) if ext in (".png", ".jpg", ".jpeg", ".webp"): return describe_image(path, question) return f"(type de fichier non géré : {ext})" def load_and_extract(state: AgentState): if not state.get("file_name"): return {"file_path": "", "file_content": ""} try: path = download_file(state["task_id"], state["file_name"]) content = extract(path, state["question"]) except Exception as e: path, content = "", f"(fichier indisponible : {e})" return {"file_path": path, "file_content": content[:20000]} # ============================================================================= # 3. Outils # ============================================================================= tavily = TavilyClient(api_key=os.getenv("TAVILY_API_KEY")) @tool def web_search(query: str) -> str: """Searches the web and returns the top results (title, url, content). Use it for recent facts or anything not on Wikipedia. Args: query: a short, precise search query (e.g. 'Malko Competition winners list'). """ try: res = tavily.search(query, max_results=5) except Exception as e: return f"web_search error: {e}" return "\n\n".join(f"{r['title']}\n{r['url']}\n{r['content']}" for r in res["results"]) or "No results." # --- RAG LlamaIndex ----------------------------------------------------------- Settings.embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5") Settings.node_parser = SentenceSplitter(chunk_size=384, chunk_overlap=48) Settings.llm = None _INDEX_CACHE = {} def fetch_page(source): if source.startswith("http"): html = requests.get(source, headers=HEADERS, timeout=30).text return source, markdownify(html, heading_style="ATX") api = "https://en.wikipedia.org/w/api.php" p = requests.get(api, headers=HEADERS, timeout=30, params={ "action": "parse", "page": source, "prop": "text", "format": "json", "redirects": 1}).json() if "error" in p: hits = requests.get(api, headers=HEADERS, timeout=30, params={ "action": "query", "list": "search", "srsearch": source, "format": "json", "srlimit": 1}).json()["query"]["search"] if not hits: raise ValueError(f"Aucune page Wikipedia pour '{source}'") return fetch_page(hits[0]["title"]) return p["parse"]["title"], markdownify(p["parse"]["text"]["*"], heading_style="ATX") def clean_md(t): t = re.sub(r"\[([^\]]*)\]\([^)]*\)", r"\1", t) t = re.sub(r"\[\[?edit\]?\]", "", t) t = re.sub(r"[ \t]+", " ", t) return re.sub(r"\n\s*\n+", "\n\n", t) def to_section_docs(page_title, md): """Un Document par section, avec son chemin de titres en métadonnée.""" docs, heads, buf = [], {}, [] def flush(): text = "\n".join(buf).strip() if text: section = " > ".join(heads[k] for k in sorted(heads)) or "Introduction" docs.append(Document(text=text, metadata={"page": page_title, "section": section})) for line in md.split("\n"): m = re.match(r"^(#{1,6})\s+(.*)", line) if m: flush() buf = [] level = len(m.group(1)) heads = {k: v for k, v in heads.items() if k < level} heads[level] = m.group(2).strip() else: buf.append(line) flush() return docs @tool def wiki_rag(source: str, query: str) -> str: """Reads a long web page or English Wikipedia page and returns the most relevant passages, each labelled with its section title (e.g. 'Discography > Studio albums'). Use a precise query (include years, names). If the passages don't answer, change the query words. Args: source: a Wikipedia page title (e.g. 'Mercedes Sosa') or a full URL. query: what you are looking for (e.g. 'studio albums discography 2000 2009'). """ try: if source not in _INDEX_CACHE: title, md = fetch_page(source) docs = to_section_docs(title, clean_md(md)) _INDEX_CACHE[source] = (title, VectorStoreIndex.from_documents(docs)) title, index = _INDEX_CACHE[source] engine = index.as_query_engine(similarity_top_k=3, response_mode="no_text") nodes = engine.query(query).source_nodes return f"Page: {title}\n\n" + "\n\n".join( f"[Section: {n.node.metadata['section']}]\n{n.node.get_content()}" for n in nodes) except Exception as e: return f"wiki_rag error: {e}" @tool("python") def python_exec(code: str) -> str: """Runs Python code and returns what it prints. Use it for ANY calculation, counting, date arithmetic, text manipulation (e.g. reversing a string) or reading files with pandas. Always print() the result you need. Args: code: complete Python code to run, with print() for the results. """ try: r = subprocess.run([sys.executable, "-c", code], capture_output=True, text=True, timeout=60) except subprocess.TimeoutExpired: return "Error: code took more than 60 seconds." out = (r.stdout + ("\nERROR:\n" + r.stderr if r.returncode else "")).strip() return out[:5000] or "(no output — did you forget print()?)" @tool def youtube_transcript(url: str) -> str: """Returns the transcript (subtitles) of a YouTube video, if available. Args: url: the YouTube video url. """ m = re.search(r"(?:v=|youtu\.be/|shorts/)([\w-]{11})", url) if not m: return "Invalid YouTube url." try: from youtube_transcript_api import YouTubeTranscriptApi parts = YouTubeTranscriptApi().fetch(m.group(1)) return " ".join(p.text for p in parts)[:15000] except Exception as e: return f"No transcript available ({type(e).__name__}). Try web_search about this video instead." tools = [web_search, wiki_rag, python_exec, youtube_transcript] TOOL_NAMES = [t.name for t in tools] # ============================================================================= # 4. Nœud assistant # ============================================================================= GAIA_PROMPT = """You are a general AI assistant. I will ask you a question. Report your thoughts, and finish your answer with the following template: FINAL ANSWER: [YOUR FINAL ANSWER] YOUR FINAL ANSWER should be a number OR as few words as possible OR a comma separated list of numbers and/or strings. If you are asked for a number, don't use comma to write your number neither use units such as $ or percent sign unless specified otherwise. If you are asked for a string, don't use articles, neither abbreviations (e.g. for cities), and write the digits in plain text unless specified otherwise. If you are asked for a comma separated list, apply the above rules depending of whether the element to be put in the list is a number or a string. Always follow any format given in the question (order, decimals, units, first name only...).""" SYSTEM_PROMPT = GAIA_PROMPT + """ Tools you can use: - wiki_rag(source, query): read a Wikipedia page or URL and get the relevant passages. Prefer it for facts on Wikipedia. - web_search(query): search the web, then use wiki_rag on a good URL to read details. - python(code): for ANY calculation, counting, sorting or text manipulation. Read attached files with pandas using the local path. - youtube_transcript(url): subtitles of a YouTube video. Check facts with tools instead of guessing. If the question looks strange (e.g. reversed text), decode it first with python.""" llm = ChatGroq(model=MODEL_ID, temperature=0, max_tokens=2000, max_retries=5) llm_with_tools = llm.bind_tools(tools) llm_final = llm.bind_tools(tools, tool_choice="none") MAX_TOOL_ROUNDS = 6 def clean_text(t): t = re.sub(r"\[([^\]]*)\]\([^)]*\)", r"\1", t) t = re.sub(r"\[\[?edit\]?\]", "", t) t = re.sub(r"[ \t]+", " ", t) t = re.sub(r"\n\s*\n+", "\n", t) return t.strip() def compact(messages, keep_full=3, full_len=4000, old_len=800): """Copie allégée de l'historique pour rester sous la limite de tokens de Groq.""" tool_idx = [i for i, m in enumerate(messages) if isinstance(m, ToolMessage)] recent = set(tool_idx[-keep_full:]) out = [] for i, m in enumerate(messages): if isinstance(m, ToolMessage): limit = full_len if i in recent else old_len m = ToolMessage(content=clean_text(str(m.content))[:limit], tool_call_id=m.tool_call_id, name=m.name) out.append(m) return out def build_user_message(state: AgentState): text = f"Question:\n{state['question']}" if state.get("file_content"): text += (f"\n\nAttached file '{state['file_name']}' (local path: {state['file_path']}). " f"Content:\n{state['file_content'][:6000]}") return HumanMessage(content=text) def repair_tool_call(failed: str): """Transforme un appel d'outil mal formé (code brut) en vrai appel d'outil.""" m = re.search(r'"name"\s*:\s*"([\w.]+)"\s*,\s*"arguments"\s*:\s*(.*)', failed, re.DOTALL) if not m: return None name, raw = m.group(1), m.group(2).strip() if raw.endswith("}"): raw = raw[:-1].rstrip() try: args = json.loads(raw) except Exception: if name != "python": return None args = {"code": raw} if name == "python" and "code" not in args: args = {"code": next(iter(args.values()), "")} if args else {"code": raw} return AIMessage(content="", tool_calls=[{"name": name, "args": args, "id": "fix_" + uuid.uuid4().hex[:8]}]) def call_llm(model, messages, tries=3): for _ in range(tries): try: return model.invoke(messages) except Exception as e: msg = str(e) body = getattr(e, "body", None) or {} failed = (body.get("error") or {}).get("failed_generation", "") if isinstance(body, dict) else "" if failed: fixed = repair_tool_call(failed) if fixed and fixed.tool_calls[0]["name"] in TOOL_NAMES: return fixed if "400" in msg and "tool" in msg.lower(): messages = messages + [HumanMessage(content=( f"Error: invalid tool call. You can ONLY use these tools: {', '.join(TOOL_NAMES)}, " f"with JSON arguments, e.g. python(code=\"print(1)\"). Try again."))] continue raise return model.invoke(messages) def get_reasoning(msg): return str(msg.additional_kwargs.get("reasoning_content", "") or "") def assistant(state: AgentState): new = [] if state["messages"] else [build_user_message(state)] history = compact(state["messages"] + new) rounds = sum(1 for m in history if getattr(m, "tool_calls", None)) model = llm_final if rounds >= MAX_TOOL_ROUNDS else llm_with_tools if rounds >= MAX_TOOL_ROUNDS: history.append(HumanMessage(content="Stop using tools. Using only the information above, " "give your best answer now, ending with 'FINAL ANSWER: ...'.")) response = call_llm(model, [SystemMessage(content=SYSTEM_PROMPT)] + history) if not response.tool_calls and not str(response.content).strip(): reasoning = get_reasoning(response) if "FINAL ANSWER" in reasoning.upper(): response = AIMessage(content=reasoning) else: retry = history + [HumanMessage(content="Your last reply was empty. " "Give your answer now, ending with 'FINAL ANSWER: ...'.")] response = call_llm(llm_final, [SystemMessage(content=SYSTEM_PROMPT)] + retry) if not str(response.content).strip(): response = AIMessage(content=get_reasoning(response)) return {"messages": new + [response]} # ============================================================================= # 5. Nœud format_answer # ============================================================================= def clean(raw: str) -> str: m = re.search(r"FINAL ANSWER\s*:\s*(.+)", raw, re.IGNORECASE | re.DOTALL) ans = m.group(1) if m else raw ans = ans.strip().split("\n")[0] ans = ans.replace("**", "").strip().strip("\"'`[]").strip() if ans.endswith("."): ans = ans[:-1] ans = re.sub(r"\s*,\s*", ", ", ans) return ans def format_answer(state: AgentState): return {"answer": clean(str(state["messages"][-1].content))} # ============================================================================= # 6. Graphe # ============================================================================= builder = StateGraph(AgentState) builder.add_node("load_and_extract", load_and_extract) builder.add_node("assistant", assistant) builder.add_node("tools", ToolNode(tools)) builder.add_node("format_answer", format_answer) builder.add_edge(START, "load_and_extract") builder.add_edge("load_and_extract", "assistant") builder.add_conditional_edges("assistant", tools_condition, {"tools": "tools", END: "format_answer"}) builder.add_edge("tools", "assistant") builder.add_edge("format_answer", END) graph = builder.compile() class GaiaAgent: """Interface simple utilisée par app.py.""" def __call__(self, question: str, task_id: str = "", file_name: str = "") -> str: state = graph.invoke( {"task_id": task_id, "question": question, "file_name": file_name or "", "messages": []}, {"recursion_limit": 40}, ) return state["answer"]