Course_Check / agent.py
SlipStream90
Switch Groq default to groq/compound (user edit)
dc9a2d8
Raw History Blame Contribute Delete
6.51 kB
"""LangGraph ReAct agent for the GAIA final assignment.
Builds a tool-calling agent graph (langgraph.prebuilt.create_react_agent)
with automatic LLM provider fallback:
1. GROQ_API_KEY -> Groq (free tier, e.g. llama-3.1-8b-instant) [preferred if HF credits depleted]
2. GOOGLE_API_KEY / GEMINI_API_KEY -> Gemini 2.0 Flash (free tier)
3. HF_TOKEN -> Hugging Face Inference Providers (Qwen2.5-72B)
Set one of these as a Space Secret. The agent picks the first available.
Optional overrides: GROQ_MODEL, GOOGLE_MODEL, HF_MODEL, HF_VLM_MODEL.
"""
import os
import re
import time
from langchain_core.messages import HumanMessage, SystemMessage
from langgraph.prebuilt import create_react_agent
from tools import ALL_TOOLS
DEFAULT_HF_MODEL = "Qwen/Qwen3-32B"
DEFAULT_GROQ_MODEL = "groq/compound"
DEFAULT_GEMINI_MODEL = "gemini-2.0-flash"
MAX_NEW_TOKENS = 512
RECURSION_LIMIT = 35
GAIA_SYSTEM_PROMPT = (
"You solve GAIA questions scored by EXACT MATCH. Use tools for every fact; never guess.\n"
"Prefer wikipedia_lookup, web_search+visit_webpage, analyze_file, calculate, run_python.\n"
"Finish by outputting ONLY the answer, no preamble, no 'FINAL ANSWER', no quotes.\n"
"Numbers: digits only, no units/commas unless asked. Keep proper-noun casing."
)
def _build_llm():
# 1) Groq (free, fast) β€” set GROQ_API_KEY secret
groq_key = os.getenv("GROQ_API_KEY")
if groq_key:
try:
from langchain_groq import ChatGroq
model = os.getenv("GROQ_MODEL", DEFAULT_GROQ_MODEL)
print(f"Using Groq model: {model}")
return ChatGroq(
model=model,
api_key=groq_key,
temperature=0.1,
max_tokens=MAX_NEW_TOKENS,
)
except Exception as e:
print(f"Groq init failed, falling back: {e}")
# 2) Google Gemini (free tier) β€” set GOOGLE_API_KEY or GEMINI_API_KEY
google_key = os.getenv("GOOGLE_API_KEY") or os.getenv("GEMINI_API_KEY")
if google_key:
try:
from langchain_google_genai import ChatGoogleGenerativeAI
model = os.getenv("GOOGLE_MODEL", DEFAULT_GEMINI_MODEL)
print(f"Using Gemini model: {model}")
return ChatGoogleGenerativeAI(
model=model,
google_api_key=google_key,
temperature=0.1,
max_output_tokens=MAX_NEW_TOKENS,
)
except Exception as e:
print(f"Gemini init failed, falling back: {e}")
# 3) Hugging Face Inference Providers
token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
if token:
try:
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
model_id = os.getenv("HF_MODEL", DEFAULT_HF_MODEL)
print(f"Using HF model: {model_id}")
endpoint = HuggingFaceEndpoint(
repo_id=model_id,
task="conversational",
provider="auto",
huggingfacehub_api_token=token,
max_new_tokens=MAX_NEW_TOKENS,
temperature=0.1,
timeout=120,
)
return ChatHuggingFace(llm=endpoint)
except Exception as e:
print(f"HF init failed: {e}")
raise RuntimeError(
"No LLM credentials found. Set one of these Space Secrets:\n"
"- GROQ_API_KEY (recommended free): get at https://console.groq.com/keys\n"
"- GOOGLE_API_KEY (Gemini free): get at https://aistudio.google.com/app/apikey\n"
"- HF_TOKEN (HF Inference Providers): https://huggingface.co/settings/tokens\n"
"After adding, restart the Space."
)
def build_gaia_graph():
llm = _build_llm()
return create_react_agent(
model=llm,
tools=ALL_TOOLS,
prompt=SystemMessage(content=GAIA_SYSTEM_PROMPT),
)
def clean_answer(raw: str) -> str:
"""Normalize the model's final message into a bare exact-match answer."""
text = raw.strip()
# Drop any FINAL ANSWER-style prefix the model may still emit.
text = re.sub(r"^final\s+answer\s*[:\-]?\s*", "", text, flags=re.I)
# Collapse internal newlines so the answer is a single line.
text = re.sub(r"\s*\n+\s*", " ", text).strip()
# Strip one layer of wrapping quotes.
if len(text) >= 2 and text[0] == text[-1] and text[0] in "\"'":
text = text[1:-1].strip()
return text
class GaiaAgent:
"""Same call signature as the template's BasicAgent, backed by LangGraph."""
def __init__(self):
self.graph = build_gaia_graph()
print("GaiaAgent initialized.")
def __call__(self, question: str, file_path: str | None = None) -> str:
content = question.strip()
if file_path:
content += (
f"\n\nA file relevant to this question has been downloaded "
f"to the local path: {file_path}\n"
f"Start by calling analyze_file on this path."
)
# Retry on 413/rate-limit (Groq free tier 8000 TPM) with backoff + shorter history
for attempt in range(3):
try:
result = self.graph.invoke(
{"messages": [HumanMessage(content=content)]},
config={"recursion_limit": RECURSION_LIMIT},
)
raw = result["messages"][-1].content
if isinstance(raw, list):
raw = " ".join(
b.get("text", "") for b in raw if isinstance(b, dict)
)
answer = clean_answer(str(raw))
print(f"Agent answer: {answer[:100]}")
return answer
except Exception as e:
msg = str(e)
is_rate = "413" in msg or "rate_limit" in msg.lower() or "RateLimit" in msg or "tokens per minute" in msg.lower()
if is_rate and attempt < 2:
wait = 12 * (attempt + 1)
print(f"Rate limited (attempt {attempt+1}/3): {msg[:200]} β€” waiting {wait}s then retrying with trimmed context")
time.sleep(wait)
# On retry, shorten the question a bit to reduce tokens
if len(content) > 2000:
content = content[:2000] + "..."
continue
# Non-rate-limit or final failure: surface as AGENT ERROR but don't crash the whole run
raise