Update app.py

app.py

@@ -1,11 +1,13 @@
 import os
 import json
 import re
+import time
+
 import gradio as gr
 import requests
 import spaces
 
-from huggingface_hub import InferenceClient
+from groq import Groq
 
 from prompt import SYSTEM_PROMPT
 from tools import execute_tool, download_file
@@ -17,29 +19,84 @@ from tools import execute_tool, download_file
 
 SCORING_API = "https://agents-course-unit4-scoring.hf.space"
 
-MODEL_ID = "Qwen/Qwen2.5-72B-Instruct"
+GROQ_API_KEY = os.getenv("GROQ_API_KEY")
+
+TEXT_MODEL = os.getenv(
+    "GROQ_TEXT_MODEL",
+    "openai/gpt-oss-120b"
+)
 
-HF_TOKEN = os.getenv("HF_TOKEN")
+MAX_STEPS = 5
 
-client = InferenceClient(
-    model=MODEL_ID,
-    token=HF_TOKEN
+client = Groq(
+    api_key=GROQ_API_KEY
 )
 
 
 # ============================================================
-# ZERO GPU STARTUP PROBE
+# ZERO GPU COMPATIBILITY
 # ============================================================
 
-# ZeroGPU requires at least one @spaces.GPU function to exist
-# when the Space starts.
+# The Space is running on ZeroGPU.
+# Our actual agent does NOT need the GPU because Groq
+# performs the model inference remotely.
 #
-# We DO NOT call this function.
-# Therefore it consumes no GPU quota.
+# This function exists so the Space is recognized as
+# ZeroGPU-compatible.
 
 @spaces.GPU(duration=1)
 def zerogpu_probe():
-    return "ok"
+    return "ZeroGPU available"
+
+
+# ============================================================
+# JSON PARSING
+# ============================================================
+
+def extract_json(text):
+
+    if not text:
+        return None
+
+    text = text.strip()
+
+    try:
+        return json.loads(text)
+    except Exception:
+        pass
+
+    # Remove markdown code fences
+    cleaned = re.sub(
+        r"```(?:json)?",
+        "",
+        text,
+        flags=re.IGNORECASE
+    )
+
+    cleaned = cleaned.replace(
+        "```",
+        ""
+    ).strip()
+
+    try:
+        return json.loads(cleaned)
+    except Exception:
+        pass
+
+    # Find first JSON object
+    start = cleaned.find("{")
+    end = cleaned.rfind("}")
+
+    if start != -1 and end != -1 and end > start:
+
+        candidate = cleaned[start:end + 1]
+
+        try:
+            return json.loads(candidate)
+        except Exception:
+            pass
+
+    return None
 
 
 # ============================================================
@@ -47,68 +104,93 @@ def zerogpu_probe():
 # ============================================================
 
 def clean_answer(answer):
+
     if answer is None:
         return ""
 
     answer = str(answer).strip()
 
-    answer = answer.replace("```", "")
-    answer = answer.replace("FINAL ANSWER:", "")
-    answer = answer.replace("Final Answer:", "")
-    answer = answer.replace("THE ANSWER IS:", "")
-    answer = answer.replace("The answer is:", "")
+    answer = re.sub(
+        r"^```(?:text|json)?\s*",
+        "",
+        answer,
+        flags=re.IGNORECASE
+    )
+
+    answer = re.sub(
+        r"\s*```$",
+        "",
+        answer
+    )
+
+    prefixes = [
+        "FINAL ANSWER:",
+        "Final Answer:",
+        "THE ANSWER IS:",
+        "The answer is:",
+        "ANSWER:"
+    ]
 
-    answer = answer.strip()
+    for prefix in prefixes:
 
-    return answer
+        if answer.startswith(prefix):
+
+            answer = answer[
+                len(prefix):
+            ].strip()
+
+    return answer.strip()
 
 
 # ============================================================
-# JSON EXTRACTION
+# GROQ CALL
 # ============================================================
 
-def extract_json(text):
+def call_groq(messages):
 
-    text = text.strip()
+    last_error = None
 
-    # Direct JSON
-    try:
-        return json.loads(text)
-    except:
-        pass
+    for attempt in range(3):
 
-    # Find JSON object inside model response
-    match = re.search(r"\{.*\}", text, re.DOTALL)
-
-    if match:
         try:
-            return json.loads(match.group(0))
-        except:
-            pass
 
-    return None
+            response = client.chat.completions.create(
+                model=TEXT_MODEL,
+                messages=messages,
+                temperature=0.1,
+                max_tokens=1200
+            )
 
+            return response.choices[0].message.content
 
-# ============================================================
-# MODEL CALL
-# ============================================================
+        except Exception as e:
 
-def ask_model(messages):
+            last_error = e
 
-    response = client.chat_completion(
-        messages=messages,
-        max_tokens=1200,
-        temperature=0.1
-    )
+            print(
+                f"Groq error "
+                f"(attempt {attempt + 1}/3): {e}"
+            )
 
-    return response.choices[0].message.content
+            # Backoff for rate limits
+            if attempt < 2:
+                time.sleep(
+                    2 ** attempt
+                )
+
+    raise RuntimeError(
+        f"Groq request failed: {last_error}"
+    )
 
 
 # ============================================================
-# SOLVE ONE GAIA QUESTION
+# SOLVE ONE QUESTION
 # ============================================================
 
-def solve_question(task_id, question):
+def solve_question(
+    task_id,
+    question
+):
 
     messages = [
         {
@@ -118,29 +200,50 @@ def solve_question(task_id, question):
         {
             "role": "user",
             "content": (
-                f"GAIA TASK ID: {task_id}\n\n"
+                f"GAIA TASK ID:\n{task_id}\n\n"
                 f"QUESTION:\n{question}\n\n"
-                "Solve this task completely. "
-                "Use tools whenever necessary."
+                "Solve this question. "
+                "Use tools only when necessary. "
+                "Return the exact requested answer."
             )
         }
     ]
 
     current_file = None
 
-    for step in range(12):
+    for step in range(MAX_STEPS):
+
+        print(
+            f"  Agent step "
+            f"{step + 1}/{MAX_STEPS}"
+        )
 
         try:
-            response = ask_model(messages)
+
+            response = call_groq(
+                messages
+            )
 
         except Exception as e:
-            return f"ERROR: model request failed: {e}"
 
-        parsed = extract_json(response)
+            return (
+                f"ERROR: model request failed: {e}"
+            )
+
+        print(
+            f"  Model: {response[:500]}"
+        )
+
+        parsed = extract_json(
+            response
+        )
+
+        # ----------------------------------------------------
+        # Model failed JSON formatting
+        # ----------------------------------------------------
 
         if parsed is None:
 
-            # Give model another chance to follow JSON format
             messages.append(
                 {
                     "role": "assistant",
@@ -152,10 +255,11 @@ def solve_question(task_id, question):
                 {
                     "role": "user",
                     "content": (
-                        "Your previous response was not valid JSON. "
-                        "Respond ONLY with either:\n"
-                        '{"action":"tool","tool":"TOOL_NAME","input":"..."}\n'
-                        "or:\n"
+                        "Invalid format. "
+                        "Respond ONLY with valid JSON.\n\n"
+                        "For a tool:\n"
+                        '{"action":"tool","tool":"TOOL_NAME","input":"..."}\n\n'
+                        "For the final answer:\n"
                         '{"action":"final","answer":"..."}'
                     )
                 }
@@ -163,60 +267,59 @@ def solve_question(task_id, question):
 
             continue
 
-        action = parsed.get("action")
+        action = parsed.get(
+            "action"
+        )
 
         # ----------------------------------------------------
-        # FINAL ANSWER
+        # FINAL
         # ----------------------------------------------------
 
         if action == "final":
 
-            answer = parsed.get("answer", "")
+            answer = parsed.get(
+                "answer",
+                ""
+            )
 
-            return clean_answer(answer)
+            return clean_answer(
+                answer
+            )
 
         # ----------------------------------------------------
-        # TOOL CALL
+        # TOOL
         # ----------------------------------------------------
 
         if action == "tool":
 
-            tool_name = parsed.get("tool")
-            tool_input = parsed.get("input", "")
+            tool_name = parsed.get(
+                "tool"
+            )
 
-            if not tool_name:
-                messages.append(
-                    {
-                        "role": "user",
-                        "content": "You must provide a tool name."
-                    }
-                )
-                continue
+            tool_input = parsed.get(
+                "input",
+                ""
+            )
 
-            try:
+            print(
+                f"  Tool: {tool_name}"
+            )
 
-                # ------------------------------------------------
-                # Download file
-                # ------------------------------------------------
+            try:
 
+                # Downloading a GAIA attachment
                 if tool_name == "download_file":
 
-                    file_path = download_file(
-                        task_id,
-                        tool_input
+                    current_file = download_file(
+                        task_id
                     )
 
-                    current_file = file_path
-
                     result = (
-                        f"File downloaded successfully.\n"
-                        f"FILE_PATH: {file_path}"
+                        "File downloaded successfully.\n"
+                        f"FILE_PATH: {current_file}"
                     )
 
-                # ------------------------------------------------
-                # File tools automatically use current file
-                # ------------------------------------------------
-
+                # Tools that require current file
                 elif tool_name in [
                     "inspect_file",
                     "transcribe_audio"
@@ -225,8 +328,8 @@ def solve_question(task_id, question):
                     if not current_file:
 
                         result = (
-                            "No file has been downloaded yet. "
-                            "Use download_file first."
+                            "No attachment has been downloaded. "
+                            "Call download_file first."
                         )
 
                     else:
@@ -241,8 +344,8 @@ def solve_question(task_id, question):
                     if not current_file:
 
                         result = (
-                            "No image has been downloaded yet. "
-                            "Use download_file first."
+                            "No image has been downloaded. "
+                            "Call download_file first."
                         )
 
                     else:
@@ -264,25 +367,39 @@ def solve_question(task_id, question):
 
             except Exception as e:
 
-                result = f"TOOL ERROR: {e}"
+                result = (
+                    f"TOOL ERROR: {e}"
+                )
 
-            # Give tool result back to model
+            # Add the tool call and result
+            # to the conversation.
             messages.append(
                 {
                     "role": "assistant",
-                    "content": json.dumps(parsed)
+                    "content": json.dumps(
+                        parsed
+                    )
                 }
             )
 
+            # Keep tool output reasonably short.
+            result = str(result)
+
+            if len(result) > 18000:
+
+                result = (
+                    result[:18000]
+                    + "\n[TOOL OUTPUT TRUNCATED]"
+                )
+
             messages.append(
                 {
                     "role": "user",
                     "content": (
-                        f"TOOL RESULT ({tool_name}):\n"
-                        f"{result}\n\n"
-                        "Continue solving the original question. "
-                        "Use another tool if necessary, otherwise return "
-                        "the final answer as JSON."
+                        f"TOOL RESULT:\n{result}\n\n"
+                        "Continue solving the original "
+                        "question. If you now know the "
+                        "answer, return the final JSON."
                     )
                 }
             )
@@ -297,18 +414,21 @@ def solve_question(task_id, question):
             {
                 "role": "user",
                 "content": (
-                    "Invalid action. Use exactly one of:\n"
+                    "Invalid action. Use exactly:\n"
                     '{"action":"tool","tool":"TOOL_NAME","input":"..."}\n'
+                    "or\n"
                     '{"action":"final","answer":"..."}'
                 )
             }
         )
 
-    return "ERROR: agent exceeded maximum reasoning steps."
+    return (
+        "ERROR: maximum reasoning steps reached."
+    )
 
 
 # ============================================================
-# GET QUESTIONS
+# GET GAIA QUESTIONS
 # ============================================================
 
 def get_questions():
@@ -334,17 +454,22 @@ def get_questions():
 
 
 # ============================================================
-# SUBMIT ANSWERS
+# SUBMIT
 # ============================================================
 
-def submit_answers(username, answers):
+def submit_answers(
+    username,
+    answers
+):
 
-    space_id = os.getenv("SPACE_ID")
+    space_id = os.getenv(
+        "SPACE_ID"
+    )
 
     if not space_id:
 
         raise RuntimeError(
-            "SPACE_ID environment variable is missing."
+            "SPACE_ID is missing."
         )
 
     agent_code = (
@@ -358,6 +483,14 @@ def submit_answers(username, answers):
         "answers": answers
     }
 
+    print(
+        "\nAgent code:"
+    )
+
+    print(
+        agent_code
+    )
+
     response = requests.post(
         f"{SCORING_API}/submit",
         json=payload,
@@ -370,49 +503,93 @@ def submit_answers(username, answers):
 
 
 # ============================================================
-# MAIN EVALUATION
+# RUN ALL 20
 # ============================================================
 
 def run_agent(username):
 
-    if not username or not username.strip():
+    if not GROQ_API_KEY:
 
         return (
-            "ERROR: Please enter your Hugging Face username.",
+            "ERROR: GROQ_API_KEY is not configured.",
+            []
+        )
+
+    if not username:
+
+        return (
+            "ERROR: Enter your Hugging Face username.",
             []
         )
 
     username = username.strip()
 
-    print("\n========================================")
+    print("\n")
+    print("=" * 60)
     print("GAIA UNIT 4 AGENT")
-    print("========================================")
+    print("=" * 60)
+
+    print(
+        f"Username: {username}"
+    )
 
-    print(f"Username: {username}")
+    print(
+        f"Model: {TEXT_MODEL}"
+    )
 
-    questions = get_questions()
+    # --------------------------------------------------------
+    # Get questions
+    # --------------------------------------------------------
 
-    print(f"Questions received: {len(questions)}")
+    try:
+
+        questions = get_questions()
+
+    except Exception as e:
+
+        return (
+            f"ERROR getting questions: {e}",
+            []
+        )
+
+    print(
+        f"Questions: {len(questions)}"
+    )
 
     answers_payload = []
+
     display_rows = []
 
-    for index, question_data in enumerate(
+    # --------------------------------------------------------
+    # Sequential processing
+    # --------------------------------------------------------
+
+    for index, item in enumerate(
         questions,
         start=1
     ):
 
-        task_id = question_data.get("task_id")
+        task_id = item.get(
+            "task_id"
+        )
 
-        question = question_data.get(
+        question = item.get(
             "question",
-            question_data.get("prompt", "")
+            item.get(
+                "prompt",
+                ""
+            )
         )
 
-        print("\n----------------------------------------")
-        print(f"Question {index}/{len(questions)}")
-        print(f"Task ID: {task_id}")
-        print("----------------------------------------")
+        print("\n")
+        print("=" * 60)
+        print(
+            f"QUESTION {index}/{len(questions)}"
+        )
+        print(
+            f"TASK ID: {task_id}"
+        )
+        print("=" * 60)
 
         try:
 
@@ -423,11 +600,17 @@ def run_agent(username):
 
         except Exception as e:
 
-            answer = f"ERROR: {e}"
+            answer = (
+                f"ERROR: {e}"
+            )
 
-        answer = clean_answer(answer)
+        answer = clean_answer(
+            answer
+        )
 
-        print(f"ANSWER: {answer}")
+        print(
+            f"FINAL ANSWER: {answer}"
+        )
 
         answers_payload.append(
             {
@@ -444,9 +627,17 @@ def run_agent(username):
             ]
         )
 
-    print("\n========================================")
-    print("SUBMITTING")
-    print("========================================")
+        # Small pause between questions.
+        time.sleep(0.5)
+
+    # --------------------------------------------------------
+    # Submit once
+    # --------------------------------------------------------
+
+    print("\n")
+    print("=" * 60)
+    print("SUBMITTING ALL ANSWERS")
+    print("=" * 60)
 
     try:
 
@@ -455,42 +646,56 @@ def run_agent(username):
             answers_payload
         )
 
-        print("Submission result:")
-        print(result)
+    except Exception as e:
 
-        score = result.get(
-            "score",
-            result.get("total_score", "N/A")
+        return (
+            f"SUBMISSION ERROR: {e}",
+            display_rows
         )
 
-        correct = result.get(
-            "correct_count",
-            result.get("correct", "N/A")
-        )
+    print(
+        "SUBMISSION RESULT:"
+    )
 
-        total = result.get(
-            "total_attempted",
-            len(answers_payload)
+    print(
+        result
+    )
+
+    score = result.get(
+        "score",
+        result.get(
+            "total_score",
+            "N/A"
         )
+    )
 
-        status = (
-            f"SUBMITTED\n"
-            f"Score: {score}\n"
-            f"Correct: {correct}/{total}"
+    correct = result.get(
+        "correct_count",
+        result.get(
+            "correct",
+            "N/A"
         )
+    )
 
-        return status, display_rows
+    total = result.get(
+        "total_attempted",
+        len(answers_payload)
+    )
 
-    except Exception as e:
+    status = (
+        "SUBMITTED\n"
+        f"Score: {score}\n"
+        f"Correct: {correct}/{total}"
+    )
 
-        return (
-            f"SUBMISSION ERROR: {e}",
-            display_rows
-        )
+    return (
+        status,
+        display_rows
+    )
 
 
 # ============================================================
-# SIMPLE UI
+# GRADIO UI
 # ============================================================
 
 with gr.Blocks(
@@ -499,22 +704,21 @@ with gr.Blocks(
 
     gr.Markdown(
         """
-        # GAIA Unit 4 Agent
+# GAIA Unit 4 Agent
 
-        Enter your Hugging Face username and run the agent.
+Groq-powered autonomous agent for the GAIA Unit 4 final assignment.
 
-        The agent will solve the 20 validation questions and submit
-        the answers automatically.
-        """
+Enter your Hugging Face username and run the evaluation.
+"""
     )
 
     username = gr.Textbox(
         label="Hugging Face Username",
-        placeholder="Your HF username"
+        placeholder="e.g. Viske"
     )
 
     run_button = gr.Button(
-        "Run Agent & Submit",
+        "Run 20 Questions & Submit",
         variant="primary"
     )
 
@@ -547,8 +751,9 @@ with gr.Blocks(
     )
 
 
-# IMPORTANT:
-# Disable Gradio SSR for better Spaces compatibility.
+# ============================================================
+# LAUNCH
+# ============================================================
 
 demo.launch(
     debug=True,