Upload 2 files

agent.py

@@ -18,29 +18,50 @@ from tools import (
     youtube_transcript,
 )
 
-DEFAULT_MODEL_ID = "meta-llama/Llama-3.3-70B-Instruct"
+# Qwen2.5-72B is a much stronger tool-using / code-writing model than Llama-3.3
+# for GAIA. Override via AGENT_MODEL_ID env var if you want to test others.
+DEFAULT_MODEL_ID = "Qwen/Qwen2.5-72B-Instruct"
 DEFAULT_PROVIDER = "auto"
 
-SYSTEM_PROMPT = """You are a careful research agent solving GAIA benchmark questions.
-
-Hard rules for the FINAL ANSWER:
-- Answers are graded by EXACT string match after light normalization. Be precise.
-- Output ONLY the answer — no prefixes, no units unless the question asks, no
-  trailing punctuation, no quotes.
-- Numbers: digits only unless the question wants words. No commas, no $, no %.
-- Strings: lowercase only if the question wants lowercase; otherwise match the
-  natural casing of the entity. Do not add articles (a, an, the).
-- Lists: comma + single space between items, in the order requested.
-
-Workflow:
-1. If the question references "the attached file", "this Excel/CSV/audio/image",
-   etc., call `download_task_file(task_id)` FIRST. The task_id is given to you.
-2. Pick the right tool: `read_table` for spreadsheets, `transcribe_audio` for
-   audio, `analyze_image` for images, `read_webpage` after `web_search` for the
-   open web, `wikipedia_search` for encyclopedic facts, `youtube_transcript`
-   for YouTube. Use Python freely for math, parsing, and logic.
-3. Cross-check facts when an answer feels uncertain. Don't fabricate.
-4. End with `final_answer("<the exact answer>")`.
+# Canonical GAIA system prompt — the exact answer-format spec the leaderboard
+# uses. Deviating from this format is the #1 cause of "right reasoning, wrong
+# score".
+SYSTEM_PROMPT = """You are a general AI assistant solving questions from the GAIA benchmark.
+
+I will ask you a question. Think step by step using the tools available. When you
+are confident, finish your answer with the following template:
+
+FINAL ANSWER: [YOUR FINAL ANSWER]
+
+YOUR FINAL ANSWER must obey these rules exactly:
+- It is a number, OR as few words as possible, OR a comma separated list of
+  numbers and/or strings.
+- If a number: digits only. No commas inside numbers. No units ($, %, 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.
 """
 
 
@@ -57,7 +78,7 @@ class GaiaAgent:
             model_id=model_id or os.getenv("AGENT_MODEL_ID", DEFAULT_MODEL_ID),
             provider=provider or os.getenv("AGENT_PROVIDER", DEFAULT_PROVIDER),
             token=token,
-            max_tokens=2048,
+            max_tokens=4096,
         )
         self.tools = [
             web_search,
@@ -72,22 +93,42 @@ class GaiaAgent:
 
     def _build_agent(self) -> CodeAgent:
         # Fresh agent per question — keeps memory clean between tasks.
-        return CodeAgent(
+        kwargs = dict(
             tools=self.tools,
             model=self.model,
-            max_steps=8,
+            max_steps=15,
             additional_authorized_imports=[
                 "pandas", "numpy", "json", "re", "math", "statistics",
                 "itertools", "collections", "datetime", "csv", "io",
-                "pathlib", "string", "base64",
+                "pathlib", "string", "base64", "urllib", "unicodedata",
             ],
         )
+        # planning_interval was added in smolagents 1.x — guard for older installs.
+        try:
+            return CodeAgent(planning_interval=4, **kwargs)
+        except TypeError:
+            return CodeAgent(**kwargs)
 
     def __call__(self, question: str, task_id: Optional[str] = None) -> str:
         agent = self._build_agent()
         prompt = SYSTEM_PROMPT
         if task_id:
             prompt += f"\nThe current task_id is: {task_id}\n"
+            # Eager-download the attachment so the agent never misses it.
+            try:
+                path = download_task_file(task_id)
+                if path and not path.startswith("NO_FILE") and not path.startswith(
+                    ("Download error", "Download failed")
+                ):
+                    prompt += (
+                        f"An attached file for this task has already been "
+                        f"downloaded to: {path}\n"
+                        f"Use the right tool on this path (read_table for "
+                        f"spreadsheets/CSV, transcribe_audio for audio, "
+                        f"analyze_image for images, or open it with Python).\n"
+                    )
+            except Exception as e:
+                print(f"  pre-download skipped: {e}")
         prompt += f"\nQuestion:\n{question}\n"
         try:
             raw = agent.run(prompt)
@@ -97,16 +138,30 @@ class GaiaAgent:
 
     @staticmethod
     def _normalize(text: str) -> str:
-        """Clean common LLM cruft so the grader's exact match has a chance."""
+        """Strip every common LLM cruft pattern so exact-match has a chance."""
         if not text:
             return text
-        t = text.strip()
-        # Strip "FINAL ANSWER:" prefix if the model left it in.
-        t = re.sub(r"^\s*final answer\s*[:\-]\s*", "", t, flags=re.IGNORECASE)
-        # Strip wrapping quotes / backticks.
-        if len(t) >= 2 and t[0] == t[-1] and t[0] in "\"'`":
+        t = 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()
+            # Trim at first newline / closing markdown.
+            t = t.split("\n")[0].strip()
+
+        # 2) Strip wrapping markdown bold/italic/code.
+        t = re.sub(r"^[*_`]+|[*_`]+$", "", t).strip()
+
+        # 3) Strip wrapping quotes.
+        if len(t) >= 2 and t[0] == t[-1] and t[0] in "\"'`“”‘’":
             t = t[1:-1].strip()
-        # Drop trailing sentence-ending punctuation unless the answer is a sentence.
-        if not re.search(r"\s", t) or t.count(" ") < 6:
+
+        # 4) Drop a single trailing period if the answer is short (not a sentence).
+        if t.count(" ") < 6:
             t = t.rstrip(".")
-        return t.strip()
+
+        # 5) Collapse internal whitespace.
+        t = re.sub(r"[ \t]+", " ", t).strip()
+
+        return t

tools.py

@@ -27,7 +27,7 @@ USER_AGENT = (
 # Web search
 # ---------------------------------------------------------------------------
 @tool
-def web_search(query: str, num_results: int = 8) -> str:
+def web_search(query: str, num_results: int = 10) -> str:
     """Search the web with Serper (Google results) and return the top hits.
 
     Args:
@@ -92,7 +92,7 @@ def web_search(query: str, num_results: int = 8) -> str:
 # Web page reader
 # ---------------------------------------------------------------------------
 @tool
-def read_webpage(url: str, max_chars: int = 8000) -> str:
+def read_webpage(url: str, max_chars: int = 15000) -> str:
     """Fetch a URL and return its main text content as Markdown.
 
     Args: