import os import re import sys import time import concurrent.futures # Force UTF-8 output on Windows to avoid charmap crashes with Unicode characters if sys.platform == "win32": sys.stdout.reconfigure(encoding="utf-8", errors="replace") sys.stderr.reconfigure(encoding="utf-8", errors="replace") import gradio as gr import requests import pandas as pd from typing import Literal, TypedDict, get_args from langchain_core.messages import HumanMessage, SystemMessage from langchain_openai import ChatOpenAI from langgraph.graph import END, StateGraph from config import DEFAULT_API_URL, HF_TOKEN, GROQ_API_KEY, OPENROUTER_API_KEY, get_prompt from tools import ( web_search, wikipedia_search, visit_webpage, get_youtube_transcript, describe_image, transcribe_audio, run_python_file, read_task_file, ) # --------------------------------------------------------------------------- # Model fallback chain (primary → backup → last-resort) # --------------------------------------------------------------------------- # Use OpenRouter for the main reasoning model (better quality) and Groq for routing (fast) GROQ_MODELS = [ {"model_id": "llama-3.3-70b-versatile"}, {"model_id": "llama-3.1-8b-instant"}, ] OPENROUTER_MODELS = [ {"model_id": "google/gemini-2.0-flash-001"}, {"model_id": "qwen/qwen-2.5-72b-instruct"}, {"model_id": "meta-llama/llama-3.3-70b-instruct"}, ] _LABELS = Literal[ "python_script", "image", "audio", "other_ext", "youtube", "research", "logic" ] def _download_task_file(task_id: str, api_url: str = DEFAULT_API_URL) -> tuple[bytes, str]: """Download a file attached to a GAIA task.""" url = f"{api_url}/files/{task_id}" try: headers = {"Authorization": f"Bearer {HF_TOKEN}"} resp = requests.get(url, headers=headers, timeout=30) except requests.exceptions.RequestException as e: print(f"[DEBUG] Download error for {task_id}: {e}") return b"", "" if resp.status_code != 200: print(f"[DEBUG] GET {url} → {resp.status_code}") return b"", "" ctype = resp.headers.get("content-type", "").lower() print(f"[DEBUG] Downloaded file for {task_id}: {len(resp.content)} bytes, type={ctype}") return resp.content, ctype class AgentState(TypedDict): question: str label: str context: str answer: str task_id: str | None file_name: str | None MAX_WORKERS = 1 # sequential to stay within rate limits QUESTION_TIMEOUT = 300 # seconds before a single question is abandoned _exhausted_models: set[str] = set() # --------------------------------------------------------------------------- # # NODES (LangGraph functions) # # --------------------------------------------------------------------------- # # Router uses Groq (fast, cheap) _llm_router = ChatOpenAI( model=GROQ_MODELS[0]["model_id"], base_url="https://api.groq.com/openai/v1", api_key=GROQ_API_KEY, timeout=60, ) # Reasoning uses OpenRouter (higher quality) _llm_answer = ChatOpenAI( model=OPENROUTER_MODELS[0]["model_id"], base_url="https://openrouter.ai/api/v1", api_key=OPENROUTER_API_KEY, timeout=120, ) def route_question(state: AgentState) -> AgentState: """Label the task so we know which toolchain to invoke.""" question = state["question"] label_values = set(get_args(_LABELS)) prompt = get_prompt( prompt_key="router", question=question, labels=", ".join(repr(v) for v in label_values), ) resp = _llm_router.invoke(prompt).content.strip().lower() state["label"] = resp if resp in label_values else "logic" return state def call_tools(state: AgentState) -> AgentState: question, label, task_id = state["question"], state["label"], state["task_id"] file_name = state.get("file_name") or "" matched_obj = re.search(r"https?://\S+", question) # ---- attachment: try download when task has a file or label suggests one ----- should_try_file = bool(task_id and file_name) if not should_try_file and task_id and label in ("python_script", "image", "audio", "other_ext"): should_try_file = True # label says there's a file — try anyway if should_try_file: blob, ctype = _download_task_file(api_url=DEFAULT_API_URL, task_id=task_id) if blob: print(f"[DEBUG] attachment type={ctype}, size={len(blob)} bytes") if "python" in ctype or file_name.endswith(".py") or (label == "python_script" and "text" in ctype): print("[DEBUG] Working with a Python attachment file") state["answer"] = run_python_file.invoke({"code": blob.decode("utf-8", errors="replace")}) state["label"] = "python_script" return state if "audio" in ctype or any(file_name.endswith(ext) for ext in [".mp3", ".wav", ".m4a", ".flac"]) or (label == "audio" and "octet" in ctype): print("[DEBUG] Working with an audio attachment file") state["context"] = transcribe_audio.invoke({"audio_bytes": blob}) state["label"] = "audio" return state if "image" in ctype or any(file_name.endswith(ext) for ext in [".png", ".jpg", ".jpeg", ".gif", ".webp"]) or (label == "image" and "octet" in ctype): print("[DEBUG] Working with an image attachment file") state["answer"] = describe_image.invoke({"img_bytes": blob, "question": question}) state["label"] = "image" return state # Excel / CSV / PDF / other binary print("[DEBUG] Working with a data file attachment") state["context"] = read_task_file.invoke({"xls_bytes": blob}) state["label"] = "other_ext" return state # ---- label-based routing (when no file was fetched) ---------- if label == "youtube": print("[TOOL] youtube_transcript") if matched_obj: url = re.sub(r'[.,;:!?")\]]+$', '', matched_obj.group(0)) print(f"[TOOL] fetching transcript for: {url}") transcript = get_youtube_transcript.invoke({"video_url": url}) if transcript and transcript != "TRANSCRIPT_UNAVAILABLE": state["context"] = transcript else: # Fallback: search web using the actual question (more informative than URL alone) print("[TOOL] Transcript unavailable — searching web for video content") search_json = web_search.invoke({"query": question[:150]}) search_json2 = web_search.invoke({"query": f"youtube video {url}"}) context_parts = [f"TRANSCRIPT_UNAVAILABLE for {url}."] if search_json and search_json != "No search results found.": context_parts.append(f"Question-based search:\n{search_json}") if search_json2 and search_json2 != "No search results found.": context_parts.append(f"Video search:\n{search_json2}") state["context"] = "\n\n".join(context_parts) else: print("[TOOL] youtube label but no URL found — falling back to web search") state["context"] = web_search.invoke({"query": question}) elif label in ("image", "audio", "python_script", "other_ext"): # File was expected but unavailable — fall back to web research print(f"[TOOL] File unavailable for '{label}' question — falling back to web search") search_json = web_search.invoke({"query": question[:150]}) wiki_text = wikipedia_search.invoke({"query": question[:100]}) context_parts = ["NOTE: The attached file for this question was not available. Answer based on web research."] if search_json and search_json != "No search results found.": context_parts.append(f"Web search:\n{search_json}") try: import json as _json hits = _json.loads(search_json) for hit in hits[:2]: link = hit.get("link", "") if link: page_content = visit_webpage.invoke({"url": link}) if page_content and "Could not fetch" not in page_content: context_parts.append(f"Page ({link}):\n{page_content[:12000]}") except Exception: pass if wiki_text and "No Wikipedia results found" not in wiki_text: context_parts.append(f"Wikipedia:\n{wiki_text}") state["context"] = "\n\n".join(context_parts) elif label == "research": print("[TOOL] research — multi-step search") # Step 1: Generate a focused search query search_query_prompt = ( "Write a short, precise search query (max 10 words) to answer this question. " "Include key proper nouns, dates, and specific terms. " "Output ONLY the query, nothing else.\n\nQuestion: " + question ) focused_query = _llm_router.invoke(search_query_prompt).content.strip().strip('"').strip("'") print(f"[TOOL] search query: {focused_query}") # Step 2: Run web search + Wikipedia in parallel search_json = web_search.invoke({"query": focused_query}) wiki_text = wikipedia_search.invoke({"query": focused_query}) context_parts = [] # Step 3: Visit top search result URLs to get full page content if search_json and search_json != "No search results found.": context_parts.append(f"WEB SEARCH RESULTS:\n{search_json}") try: import json as _json hits = _json.loads(search_json) # Visit top 2 result URLs for detailed content visited = 0 for hit in hits[:4]: link = hit.get("link", "") if link and visited < 2: page_content = visit_webpage.invoke({"url": link}) if page_content and "Could not fetch" not in page_content: context_parts.append(f"\nPAGE CONTENT ({link}):\n{page_content[:15000]}") visited += 1 except Exception as e: print(f"[TOOL] Error visiting search results: {e}") if wiki_text and "No Wikipedia results found" not in wiki_text and "failed" not in wiki_text.lower(): context_parts.append(f"\nWIKIPEDIA:\n{wiki_text}") # Step 4: If initial results are thin, try an alternative query if not context_parts or all("No " in p or "error" in p.lower() for p in context_parts): print("[TOOL] Initial search thin — trying alternative query") alt_query = focused_query.replace('"', '').replace("'", "") if alt_query != focused_query: alt_results = web_search.invoke({"query": alt_query}) if alt_results and alt_results != "No search results found.": context_parts.append(f"\nALTERNATIVE SEARCH:\n{alt_results}") state["context"] = "\n\n".join(context_parts) if context_parts else "No information found from web search or Wikipedia." else: # Logic / pure reasoning — no search needed print("[TOOL] reasoning only (no search)") state["context"] = "" return state def _do_research(question: str, query: str | None = None) -> str: """Run a research search and return combined context string.""" import json as _json if not query: query = question[:120] search_json = web_search.invoke({"query": query}) context_parts = [] if search_json and search_json != "No search results found.": context_parts.append(f"Search results:\n{search_json}") try: hits = _json.loads(search_json) for hit in hits[:2]: link = hit.get("link", "") if link: page_content = visit_webpage.invoke({"url": link}) if page_content and "Could not fetch" not in page_content: context_parts.append(f"Page ({link}):\n{page_content[:12000]}") except Exception: pass return "\n\n".join(context_parts) def synthesize_response(state: AgentState) -> AgentState: # If a tool produced a direct final answer (python execution), skip reasoning if state.get("answer") and state["label"] == "python_script": print(f"[SYNTHESIZE] skipped — python output: {state['answer'][:200]}") return state # For image: the vision model already answered, but wrap it through reasoning # to extract the precise answer from the description if state.get("answer") and state["label"] == "image": state["context"] = f"VISION MODEL OUTPUT:\n{state['answer']}" state["answer"] = "" # clear so reasoning runs # For other_ext with context (file data), make sure reasoning runs if state["label"] == "other_ext" and state.get("context") and not state.get("answer"): pass # context is set, reasoning will run below # Pass 1: chain-of-thought reasoning reasoning_prompt = [ SystemMessage(content=get_prompt("reasoning_system")), HumanMessage( content=get_prompt( prompt_key="reasoning_user", question=state["question"], context=state["context"], ) ), ] reasoning = _llm_answer.invoke(reasoning_prompt).content.strip() print(f"\n[REASONING]\n{reasoning}\n") # Try to extract FINAL ANSWER directly from reasoning text fa_match = re.search(r"FINAL ANSWER:\s*(.+)", reasoning, re.IGNORECASE) if fa_match: answer = fa_match.group(1).strip().split('\n')[0].strip() elif reasoning.strip(): extract_prompt = [ SystemMessage(content=get_prompt("extract_system")), HumanMessage( content=get_prompt( prompt_key="extract_user", reasoning=reasoning, ) ), ] answer = _llm_answer.invoke(extract_prompt).content.strip() else: answer = "ERROR: no reasoning produced" # --- SECOND PASS: if answer is vague/uncertain, do more research and retry --- _bad = any(w in answer.lower() for w in ["cannot", "unable", "not determine", "no answer", "not possible"]) if _bad and state["label"] in ("research", "image", "audio", "python_script", "other_ext", "youtube"): print(f"[SYNTHESIZE] Answer '{answer[:60]}' is uncertain — doing second research pass") # Generate a more targeted query from the first reasoning refine_prompt = ( "Based on this question and partial reasoning, write a very specific search query " "(max 12 words) that would find the EXACT missing fact. " "Output ONLY the query.\n\n" f"Question: {state['question']}\n" f"Partial reasoning: {reasoning[:500]}" ) try: refined_query = _llm_router.invoke(refine_prompt).content.strip().strip('"').strip("'") print(f"[TOOL] second-pass query: {refined_query}") extra_context = _do_research(state["question"], refined_query) if extra_context: combined_context = state["context"] + "\n\nADDITIONAL RESEARCH:\n" + extra_context reasoning_prompt2 = [ SystemMessage(content=get_prompt("reasoning_system")), HumanMessage( content=get_prompt( prompt_key="reasoning_user", question=state["question"], context=combined_context, ) ), ] reasoning2 = _llm_answer.invoke(reasoning_prompt2).content.strip() print(f"\n[REASONING PASS 2]\n{reasoning2}\n") fa_match2 = re.search(r"FINAL ANSWER:\s*(.+)", reasoning2, re.IGNORECASE) if fa_match2: answer2 = fa_match2.group(1).strip().split('\n')[0].strip() _still_bad = any(w in answer2.lower() for w in ["cannot", "unable", "not determine"]) if not _still_bad: answer = answer2 except Exception as e: print(f"[SYNTHESIZE] Second pass error: {e}") state["answer"] = answer return state def format_output(state: AgentState) -> AgentState: txt = re.sub(r"^(final answer:?\s*)", "", state["answer"], flags=re.I).strip() # If question demands a single token (first name / one word), enforce it if any(kw in state["question"].lower() for kw in ["first name", "single word"]): txt = txt.split(" ")[0] state["answer"] = txt.rstrip(".") print(f"[FINAL ANSWER] {state['answer']}\n" + "-" * 60) return state # --------------------------------------------------------------------------- # # BUILD THE GRAPH # # --------------------------------------------------------------------------- # def build_graph() -> StateGraph: g = StateGraph(AgentState) g.set_entry_point("route_question") g.add_node("route_question", route_question) g.add_node("invoke_tools", call_tools) g.add_node("synthesize_response", synthesize_response) g.add_node("format_output", format_output) g.add_edge("route_question", "invoke_tools") g.add_edge("invoke_tools", "synthesize_response") g.add_edge("synthesize_response", "format_output") g.add_edge("format_output", END) return g.compile() class LGAgent: """Callable wrapper used by run_and_submit_all.""" def __init__(self, model_id: str | None = None, answer_model_id: str | None = None) -> None: global _llm_router, _llm_answer # Router: fast Groq model router_mid = model_id or GROQ_MODELS[0]["model_id"] _llm_router = ChatOpenAI( model=router_mid, base_url="https://api.groq.com/openai/v1", api_key=GROQ_API_KEY, timeout=60, ) # Answering: higher quality OpenRouter model answer_mid = answer_model_id or OPENROUTER_MODELS[0]["model_id"] _llm_answer = ChatOpenAI( model=answer_mid, base_url="https://openrouter.ai/api/v1", api_key=OPENROUTER_API_KEY, timeout=120, ) self.graph = build_graph() def __call__(self, question: str, task_id: str | None = None, file_name: str | None = None) -> str: try: state: AgentState = { "question": question, "label": "general", "context": "", "answer": "", "task_id": task_id, "file_name": file_name, } final = self.graph.invoke(state) route = final["label"] print(f"[ROUTE] '{route}' | Q: {question[:80]}") return final["answer"] except Exception as e: print("Agent error:", e) msg = str(e) # Re-raise rate-limit errors so _answer_question can fall back to the next model if "rate_limit_exceeded" in msg or "429" in msg or "413" in msg or "Request too large" in msg or "model_decommissioned" in msg or "decommissioned" in msg: raise return f"AGENT ERROR: {e}" def _parse_retry_after(error_msg: str) -> float: """Parse the suggested wait time (seconds) from a Groq 429 error message.""" m = re.search(r'try again in (?:(\d+)m)?(\d+(?:\.\d+)?)s', error_msg) if m: return float(m.group(1) or 0) * 60 + float(m.group(2)) return 65.0 # safe default def _to_str(val) -> str: """Ensure the submitted answer is always a plain string.""" if isinstance(val, str): return val if isinstance(val, list): parts = [item.get("text", "") if isinstance(item, dict) else str(item) for item in val] return " ".join(parts).strip() or "ERROR: empty response" return str(val) def _answer_question(item: dict) -> str: """Instantiate a fresh agent and answer one question, retrying on errors.""" question_text = item["question"] task_id = item.get("task_id", "") file_name = item.get("file_name") or "" augmented_question = question_text # Try each OpenRouter answer model with Groq router for answer_cfg in OPENROUTER_MODELS: answer_model_id = answer_cfg["model_id"] if answer_model_id in _exhausted_models: print(f"[{answer_model_id}] Skipped (previously rate-limited)") continue for attempt in range(2): try: result = LGAgent( model_id=GROQ_MODELS[0]["model_id"], answer_model_id=answer_model_id, )(augmented_question, task_id=task_id, file_name=file_name) # Pause between questions to respect rate limits time.sleep(3) return result except Exception as e: msg = str(e) if "model_decommissioned" in msg or "decommissioned" in msg: _exhausted_models.add(answer_model_id) print(f"[{answer_model_id}] Model decommissioned — skipping permanently") break if "rate_limit_exceeded" in msg or "429" in msg or "413" in msg or "Request too large" in msg: if "on tokens per day" in msg or "TPD" in msg: _exhausted_models.add(answer_model_id) print(f"[{answer_model_id}] Daily token limit hit — skipping for remaining questions") break wait = _parse_retry_after(msg) print(f"[{answer_model_id}] Rate limited — waiting {wait:.0f}s then retry") time.sleep(min(wait, 30)) continue else: print(f"[{answer_model_id}] Error: {msg[:200]}") break # try next model return "AGENT ERROR: all models exhausted" 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 if profile: 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 questions_url = f"{api_url}/questions" submit_url = f"{api_url}/submit" # 1. Agent is instantiated per-question inside _answer_question for parallel execution agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main" print(agent_code) # 2. Fetch Questions print(f"Fetching questions from: {questions_url}") try: response = requests.get(questions_url, timeout=15) 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 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 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 (parallel) results_log = [] answers_payload = [] valid_items = [ item for item in questions_data if item.get("task_id") and item.get("question") is not None ] print(f"Running agent on {len(valid_items)} questions") with concurrent.futures.ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor: future_to_item = { executor.submit(_answer_question, item): item for item in valid_items } for future in concurrent.futures.as_completed(future_to_item): item = future_to_item[future] task_id = item["task_id"] question_text = item["question"] try: submitted_answer = _to_str(future.result(timeout=QUESTION_TIMEOUT)) except concurrent.futures.TimeoutError: print(f"Timeout on task {task_id}") submitted_answer = "TIMEOUT" except Exception as e: print(f"Error running agent on task {task_id}: {e}") submitted_answer = f"AGENT ERROR: {e}" answers_payload.append({"task_id": task_id, "submitted_answer": _to_str(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) # 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) # 5. Submit print(f"Submitting {len(answers_payload)} answers to: {submit_url}") try: response = requests.post(submit_url, json=submission_data, timeout=60) 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"Message: {result_data.get('message', 'No message received.')}" ) print("Submission successful.") results_df = pd.DataFrame(results_log) return final_status, results_df except requests.exceptions.HTTPError as e: error_detail = f"Server responded with status {e.response.status_code}." try: error_json = e.response.json() 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 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 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 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 # --- Build Gradio Interface using Blocks --- with gr.Blocks() as demo: gr.Markdown("# Basic Agent Evaluation 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. """ ) gr.LoginButton() 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 results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True) 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 space_host_startup = os.getenv("SPACE_HOST") space_id_startup = os.getenv("SPACE_ID") # Get SPACE_ID at startup 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") else: print("ℹ️ SPACE_ID environment variable not found (running locally?). Repo URL cannot be determined.") print("-"*(60 + len(" App Starting ")) + "\n") print("Launching Gradio Interface for Basic Agent Evaluation...") demo.launch(debug=True, share=False)