154 lines
5.2 KiB
Python
154 lines
5.2 KiB
Python
"""
|
||
# main.py
|
||
# Implementation of the LangGraph reflective agent using deepagents
|
||
# Author: Auto-generated for the assignment
|
||
# Requires: deepagents, langchain-openai, langgraph, langchain
|
||
"""
|
||
import os
|
||
import asyncio
|
||
from typing import TypedDict, Annotated
|
||
|
||
from langchain_openai import ChatOpenAI
|
||
from langchain_core.messages import HumanMessage, SystemMessage
|
||
from langchain.tools import tool
|
||
from deepagents import create_deep_agent
|
||
from deepagents.backends import CompositeBackend, LocalShellBackend, FilesystemBackend
|
||
from langgraph.graph import StateGraph, START, END
|
||
from langgraph.graph.message import add_messages
|
||
|
||
# ---------- LLM configuration (OpenRouter) ----------
|
||
llm = ChatOpenAI(
|
||
model="openai/gpt-oss-20b:free",
|
||
base_url="https://openrouter.ai/api/v1",
|
||
api_key=os.getenv("OPENAI_API_KEY"),
|
||
temperature=0.0,
|
||
)
|
||
|
||
# ---------- State definition ----------
|
||
class ReflectState(TypedDict):
|
||
question: str
|
||
draft: str
|
||
critique: str
|
||
verdict: str # "ok" | "needs_revision"
|
||
round: int
|
||
max_rounds: int
|
||
|
||
# ---------- Node implementations ----------
|
||
async def draft_answer(state: ReflectState) -> ReflectState:
|
||
"""Generate an initial draft answer (5–10 sentences)."""
|
||
prompt = (
|
||
"You are an expert tutor. Answer the following question in 5–10 concise sentences. "
|
||
"Avoid filler and keep it clear.
|
||
"
|
||
f"Question: {state['question']}"
|
||
)
|
||
response = await llm.ainvoke([SystemMessage(content="You are a helpful tutor."), HumanMessage(content=prompt)])
|
||
state["draft"] = response.content.strip()
|
||
state["round"] = 1
|
||
return state
|
||
|
||
async def reflect(state: ReflectState) -> ReflectState:
|
||
"""Critique the draft and decide if revision is needed."""
|
||
prompt = (
|
||
"You are a critical reviewer. Evaluate the following answer for completeness, specificity, and lack of filler. "
|
||
"Respond with a verdict of either "ok" or "needs_revision", followed by 2–3 bullet points of constructive feedback.
|
||
"
|
||
f"Answer draft:\n{state['draft']}"
|
||
)
|
||
response = await llm.ainvoke([SystemMessage(content="You are a critical reviewer."), HumanMessage(content=prompt)])
|
||
text = response.content.strip()
|
||
# Parse verdict and critique
|
||
if "needs_revision" in text.lower():
|
||
verdict = "needs_revision"
|
||
else:
|
||
verdict = "ok"
|
||
# Extract critique lines after the verdict
|
||
lines = text.splitlines()
|
||
critique_lines = [line for line in lines if line.strip() and line.strip().lower() not in {"ok", "needs_revision"}]
|
||
critique = "\n".join(critique_lines).strip()
|
||
state["critique"] = critique
|
||
state["verdict"] = verdict
|
||
return state
|
||
|
||
async def rewrite(state: ReflectState) -> ReflectState:
|
||
"""Rewrite the draft incorporating the critique."""
|
||
prompt = (
|
||
"You are revising an answer based on the following critique. Produce a new version that addresses the points and remains 5–10 sentences.
|
||
"
|
||
f"Original draft:\n{state['draft']}\n\nCritique:\n{state['critique']}"
|
||
)
|
||
response = await llm.ainvoke([SystemMessage(content="You are a revising tutor."), HumanMessage(content=prompt)])
|
||
state["draft"] = response.content.strip()
|
||
state["round"] += 1
|
||
return state
|
||
|
||
# ---------- Graph construction ----------
|
||
def build_graph() -> StateGraph[ReflectState]:
|
||
graph = StateGraph(ReflectState)
|
||
graph.add_node("draft_answer", draft_answer)
|
||
graph.add_node("reflect", reflect)
|
||
graph.add_node("rewrite", rewrite)
|
||
|
||
# Entry point
|
||
graph.set_entry_point("draft_answer")
|
||
|
||
# Transitions
|
||
graph.add_conditional_edges(
|
||
"draft_answer",
|
||
lambda _: "reflect",
|
||
)
|
||
graph.add_conditional_edges(
|
||
"reflect",
|
||
lambda state: "END" if state["verdict"] == "ok" else "rewrite",
|
||
)
|
||
graph.add_conditional_edges(
|
||
"rewrite",
|
||
lambda state: "END" if state["round"] > state["max_rounds"] else "reflect",
|
||
)
|
||
|
||
return graph
|
||
|
||
# ---------- Tool that runs the graph ----------
|
||
@tool
|
||
async def answer_question(query: str) -> str:
|
||
"""Run the reflective LangGraph to answer a question."""
|
||
graph = build_graph()
|
||
# Initialize state
|
||
state: ReflectState = {
|
||
"question": query,
|
||
"draft": "",
|
||
"critique": "",
|
||
"verdict": "",
|
||
"round": 0,
|
||
"max_rounds": 2,
|
||
}
|
||
# Run graph
|
||
final_state = await graph.ainvoke(state)
|
||
return final_state["draft"]
|
||
|
||
# ---------- DeepAgent setup ----------
|
||
backend = CompositeBackend([
|
||
LocalShellBackend(workspace_dir="./workspace"),
|
||
FilesystemBackend(),
|
||
])
|
||
|
||
agent = create_deep_agent(
|
||
model=llm,
|
||
tools=[answer_question],
|
||
backend=backend,
|
||
system_prompt="You are a helpful educational agent. Use the provided tools to answer questions.",
|
||
)
|
||
|
||
# ---------- CLI entry point ----------
|
||
async def main():
|
||
question = "Объясни студенту разницу между tool и resource в MCP"
|
||
result = await agent.ainvoke(
|
||
{"messages": [HumanMessage(content=question)]},
|
||
{"configurable": {"thread_id": "session-1"}},
|
||
)
|
||
# The tool returns the final answer as the last message content
|
||
print("\nFinal answer:\n", result["messages"][-1].content)
|
||
|
||
if __name__ == "__main__":
|
||
asyncio.run(main())
|