commit

agent.py

@@ -0,0 +1,139 @@
+# agent.py
+import os
+import pickle
+
+from langchain.tools.retriever import create_retriever_tool
+from langchain_community.document_loaders import WikipediaLoader, ArxivLoader
+from langchain_community.vectorstores import FAISS
+from langchain_huggingface import HuggingFaceEmbeddings
+
+from langchain_openai import ChatOpenAI
+from langchain_core.documents import Document
+from langchain_core.messages import SystemMessage, HumanMessage
+from langchain_core.tools import tool
+from langchain_community.tools import DuckDuckGoSearchRun
+
+from langgraph.graph import START, StateGraph, MessagesState
+from langgraph.prebuilt import ToolNode, tools_condition
+
+ddg = DuckDuckGoSearchRun()
+
+# -----------------------
+# Tools
+# -----------------------
+@tool
+def wiki_search(query: str) -> dict:
+    docs = WikipediaLoader(query=query, load_max_docs=2).load()
+    text = "\n\n".join(d.page_content for d in docs)
+    return {"wiki": text}
+
+
+@tool
+def arxiv_search(query: str) -> dict:
+    docs = ArxivLoader(query=query, load_max_docs=2).load()
+    text = "\n\n".join(d.page_content[:1000] for d in docs)
+    return {"arxiv": text}
+
+
+ddg = DuckDuckGoSearchRun()
+
+@tool
+def web_search(query: str) -> dict:
+    """Search web using DuckDuckGo (no API key required)"""
+    try:
+        result = ddg.run(query)
+        return {"web": result}
+    except Exception as e:
+        return {"web": ""}
+
+
+TOOLS = [wiki_search, arxiv_search, web_search]
+
+
+# -----------------------
+# System Prompt
+# -----------------------
+SYSTEM_PROMPT = """
+You are solving GAIA benchmark questions.
+
+You MUST:
+- Use tools if factual information is required.
+- Reason internally but DO NOT reveal reasoning.
+- Output ONLY the final answer.
+- No explanation.
+- No extra text.
+""".strip()
+
+SYS_MSG = SystemMessage(content=SYSTEM_PROMPT)
+
+
+# -----------------------
+# Retriever (FAISS 유지)
+# -----------------------
+embeddings = HuggingFaceEmbeddings(
+    model_name="sentence-transformers/all-mpnet-base-v2"
+)
+
+if os.path.exists("faiss.pkl"):
+    with open("faiss.pkl", "rb") as f:
+        vector_store = pickle.load(f)
+else:
+    seed_docs = [
+        Document(page_content="GAIA questions require factual exact answers."),
+    ]
+    vector_store = FAISS.from_documents(seed_docs, embeddings)
+    with open("faiss.pkl", "wb") as f:
+        pickle.dump(vector_store, f)
+
+retriever_tool = create_retriever_tool(
+    retriever=vector_store.as_retriever(),
+    name="question_retriever",
+    description="Retrieve similar factual questions",
+)
+
+
+# -----------------------
+# Graph Builder
+# -----------------------
+def build_agent():
+    llm = ChatOpenAI(
+        model="gpt-4o-mini",
+        temperature=0,
+        max_tokens=128,
+    )
+
+    llm_with_tools = llm.bind_tools(TOOLS)
+
+    def retriever(state: MessagesState):
+        return {"messages": [SYS_MSG] + state["messages"]}
+
+    def assistant(state: MessagesState):
+        return {"messages": [llm_with_tools.invoke(state["messages"])]}
+
+    builder = StateGraph(MessagesState)
+
+    builder.add_node("retriever", retriever)
+    builder.add_node("assistant", assistant)
+    builder.add_node("tools", ToolNode(TOOLS))
+
+    builder.add_edge(START, "retriever")
+    builder.add_edge("retriever", "assistant")
+    builder.add_conditional_edges("assistant", tools_condition)
+    builder.add_edge("tools", "assistant")
+
+    return builder.compile()
+
+
+# -----------------------
+# Public API
+# -----------------------
+class BasicAgent:
+    def __init__(self):
+        self.graph = build_agent()
+        print("✅ LangGraph GPT-4o-mini Agent initialized")
+
+    def __call__(self, question: str) -> str:
+        result = self.graph.invoke(
+            {"messages": [HumanMessage(content=question)]}
+        )
+        return result["messages"][-1].content.strip()

app.py

@@ -7,101 +7,12 @@ import inspect
 import pandas as pd
 from typing import TypedDict
 
-from langchain_openai import ChatOpenAI
-from langchain_core.messages import HumanMessage
-from langchain_community.tools import DuckDuckGoSearchRun
+from agent import BasicAgent
 
 # (Keep Constants as is)
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
-SYSTEM_PROMPT = """
-You are solving GAIA benchmark questions.
-
-You MUST:
-- Use the provided search results as the source of truth.
-- Reason internally but DO NOT show reasoning.
-- Output ONLY the final answer.
-- No explanation.
-- No extra text.
-"""
-
-def clean_answer(text: str) -> str:
-    if not text:
-        return ""
-    s = text.strip()
-    s = s.replace("Final answer:", "").replace("Answer:", "").strip()
-    s = s.splitlines()[0].strip()
-    s = s.strip('"\'`')
-    if len(s) > 1 and s.endswith("."):
-        s = s[:-1].strip()
-    return s
-
-# -------------------------------
-# State
-# -------------------------------
-class AgentState(TypedDict):
-    question: str
-    answer: str
-
-# -------------------------------
-# Tools & LLM
-# -------------------------------
-# Search tool (무료)
-search_tool = DuckDuckGoSearchRun()
-
-# LLM (OpenAI – 이미 네 환경에서 동작 확인됨)
-llm = ChatOpenAI(
-    model="gpt-4o",
-    temperature=0,
-    max_tokens=96,
-)
-
-# -------------------------------
-# Agent
-# -------------------------------
-class BasicAgent:
-    def __init__(self):
-        print("Search-based GAIA Agent initialized.")
-
-    def __call__(self, question: str) -> str:
-        print(f"Question: {question[:80]}...")
-
-        queries = [
-            question,
-            f"{question} wikipedia",
-            f"{question} site:wikipedia.org",
-            f"{question} fact",
-        ]
-
-        snippets = []
-        for q in queries:
-            try:
-                r = search_tool.run(q)
-                if r:
-                    snippets.append(r)
-                time.sleep(0.5)  # rate-limit 회피
-            except Exception as e:
-                print("Search error:", e)
-
-        search_result = "\n\n".join(snippets)
-
-        prompt = f"""
-{SYSTEM_PROMPT}
-
-Question:
-{question}
-
-Search Results:
-{search_result}
-""".strip()
-
-        response = llm.invoke([HumanMessage(content=prompt)])
-        answer = clean_answer(response.content)
-
-        print(f"Answer: {answer}")
-        return answer
-
 def run_and_submit_all( profile: gr.OAuthProfile | None):
     """
     Fetches all questions, runs the BasicAgent on them, submits all answers,

requirements.txt

@@ -1,8 +1,10 @@
 gradio
 requests
+
 langgraph
-langchain_openai
-langchain_core
+langchain-core
 langchain-community
+langchain-openai
+
 ddgs
 duckduckgo-search