fix: main.py — Повторный экзамен #2: Граф с рефлексией на код
This commit is contained in:
@@ -1,20 +1,21 @@
|
||||
import os
|
||||
import asyncio
|
||||
from typing import TypedDict, Annotated, Dict
|
||||
from typing import TypedDict, Dict
|
||||
|
||||
from dotenv import load_dotenv
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langchain_core.messages import HumanMessage
|
||||
from langchain.tools import tool
|
||||
from deepagents import create_deep_agent
|
||||
from deepagents.backends import FilesystemBackend, LocalShellBackend, CompositeBackend
|
||||
|
||||
from langgraph.graph import StateGraph, START, END
|
||||
from langgraph.graph.message import add_messages
|
||||
|
||||
from pydantic import BaseModel, Field
|
||||
from langchain_core.output_parsers import PydanticOutputParser
|
||||
from deepagents import create_deep_agent
|
||||
from deepagents.backends import FilesystemBackend
|
||||
|
||||
# ---------- LLM ----------
|
||||
# Load API key from .env
|
||||
load_dotenv()
|
||||
|
||||
# LLM configuration - OpenRouter
|
||||
llm = ChatOpenAI(
|
||||
model="openai/gpt-oss-20b:free",
|
||||
base_url="https://openrouter.ai/api/v1",
|
||||
@@ -22,124 +23,140 @@ llm = ChatOpenAI(
|
||||
temperature=0.0,
|
||||
)
|
||||
|
||||
# ---------- Backend ----------
|
||||
backend = CompositeBackend([
|
||||
LocalShellBackend(workspace_dir="./workspace"),
|
||||
FilesystemBackend(),
|
||||
])
|
||||
# Backend for deepagents - simple filesystem
|
||||
backend = FilesystemBackend()
|
||||
|
||||
# ---------- State ----------
|
||||
# Create a deepagents agent that will be used inside the graph nodes
|
||||
agent = create_deep_agent(
|
||||
model=llm,
|
||||
tools=[],
|
||||
backend=backend,
|
||||
system_prompt="You are a code review assistant.",
|
||||
)
|
||||
|
||||
# ---------- State definition ----------
|
||||
class CodeReviewState(TypedDict):
|
||||
code: str
|
||||
draft_review: str
|
||||
criteria_scores: Dict[str, int]
|
||||
weakest_criterion: str
|
||||
verdict: str
|
||||
verdict: str # "ok" | "needs_revision"
|
||||
round: int
|
||||
max_rounds: int
|
||||
|
||||
# ---------- Pydantic model for reflect output ----------
|
||||
class ReflectOutput(BaseModel):
|
||||
pep8: int = Field(description="Score for PEP8 compliance (0-10)")
|
||||
type_hints: int = Field(description="Score for type hints (0-10)")
|
||||
edge_cases: int = Field(description="Score for edge case handling (0-10)")
|
||||
naming: int = Field(description="Score for naming conventions (0-10)")
|
||||
weakest_criterion: str = Field(description="The criterion with the lowest score")
|
||||
verdict: str = Field(description="'ok' or 'needs_revision'")
|
||||
# ---------- Structured output for critic ----------
|
||||
class CriticOutput(BaseModel):
|
||||
scores: Dict[str, int] = Field(
|
||||
description="Scores for each criterion: pep8, type_hints, edge_cases, naming. Values 0-10."
|
||||
)
|
||||
verdict: str = Field(
|
||||
description='Verdict: "ok" if all scores >= 7, otherwise "needs_revision".'
|
||||
)
|
||||
|
||||
parser = PydanticOutputParser(pydantic_object=ReflectOutput)
|
||||
critic_parser = PydanticOutputParser(pydantic_object=CriticOutput)
|
||||
|
||||
# ---------- Nodes ----------
|
||||
# ---------- Graph nodes ----------
|
||||
async def draft_review(state: CodeReviewState) -> CodeReviewState:
|
||||
prompt = f"""
|
||||
Write a concise code review for the following Python function. Provide 3-6 bullet points.
|
||||
|
||||
Function:
|
||||
```python
|
||||
{state['code']}
|
||||
```
|
||||
"""
|
||||
response = await llm.ainvoke(HumanMessage(content=prompt))
|
||||
state['draft_review'] = response.content.strip()
|
||||
prompt = (
|
||||
f"Write a concise code review (3-6 bullet points) for the following Python function:\n\n"
|
||||
f"{state['code']}\n\n"
|
||||
"Focus on style, correctness, and potential improvements."
|
||||
)
|
||||
response = await agent.ainvoke(
|
||||
{"messages": [HumanMessage(content=prompt)]},
|
||||
{"configurable": {"thread_id": "draft_review"}},
|
||||
)
|
||||
review_text = response["messages"][-1].content.strip()
|
||||
state["draft_review"] = review_text
|
||||
print("\n--- Draft Review ---")
|
||||
print(review_text)
|
||||
return state
|
||||
|
||||
async def reflect(state: CodeReviewState) -> CodeReviewState:
|
||||
prompt = f"""
|
||||
Evaluate the following code and draft review. Score each of the four criteria on a scale 0-10:
|
||||
- PEP8 compliance
|
||||
- Type hints
|
||||
- Edge case handling
|
||||
- Naming conventions
|
||||
|
||||
Provide the scores, identify the weakest criterion, and give a verdict ('ok' or 'needs_revision').
|
||||
|
||||
Code:
|
||||
```python
|
||||
{state['code']}
|
||||
```
|
||||
|
||||
Draft Review:
|
||||
```text
|
||||
{state['draft_review']}
|
||||
```
|
||||
|
||||
Return a JSON object with keys: pep8, type_hints, edge_cases, naming, weakest_criterion, verdict.
|
||||
"""
|
||||
response = await llm.ainvoke(HumanMessage(content=prompt))
|
||||
prompt = (
|
||||
f"Evaluate the following code review and assign scores (0-10) for each criterion:\n\n"
|
||||
f"Review:\n{state['draft_review']}\n\n"
|
||||
"Criteria:\n"
|
||||
"1. pep8: adherence to PEP8 style guide.\n"
|
||||
"2. type_hints: presence and correctness of type hints.\n"
|
||||
"3. edge_cases: handling of edge cases and robustness.\n"
|
||||
"4. naming: clarity and consistency of names.\n\n"
|
||||
"Return a JSON object with keys 'scores' (dict) and 'verdict' ('ok' or 'needs_revision')."
|
||||
)
|
||||
response = await agent.ainvoke(
|
||||
{"messages": [HumanMessage(content=prompt)]},
|
||||
{"configurable": {"thread_id": "reflect"}},
|
||||
)
|
||||
raw_output = response["messages"][-1].content.strip()
|
||||
try:
|
||||
parsed = parser.parse(response.content)
|
||||
parsed = critic_parser.parse(raw_output)
|
||||
except Exception as e:
|
||||
# Fallback: set all scores to 0 and verdict to needs_revision
|
||||
parsed = ReflectOutput(pep8=0, type_hints=0, edge_cases=0, naming=0, weakest_criterion="unknown", verdict="needs_revision")
|
||||
state['criteria_scores'] = {
|
||||
"pep8": parsed.pep8,
|
||||
"type_hints": parsed.type_hints,
|
||||
"edge_cases": parsed.edge_cases,
|
||||
"naming": parsed.naming,
|
||||
}
|
||||
state['weakest_criterion'] = parsed.weakest_criterion
|
||||
state['verdict'] = parsed.verdict
|
||||
# Fallback: simple parsing if JSON is malformed
|
||||
import json
|
||||
parsed = CriticOutput(**json.loads(raw_output))
|
||||
state["criteria_scores"] = parsed.scores
|
||||
# Determine weakest criterion
|
||||
weakest = min(parsed.scores.items(), key=lambda kv: kv[1])[0]
|
||||
state["weakest_criterion"] = weakest
|
||||
state["verdict"] = parsed.verdict
|
||||
print("\n--- Critic Scores ---")
|
||||
for crit, score in parsed.scores.items():
|
||||
print(f"{crit}: {score}")
|
||||
print(f"Weakest criterion: {weakest}")
|
||||
print(f"Verdict: {parsed.verdict}")
|
||||
return state
|
||||
|
||||
async def rewrite(state: CodeReviewState) -> CodeReviewState:
|
||||
prompt = f"""
|
||||
Rewrite the section of the draft review that addresses the weakest criterion: {state['weakest_criterion']}.
|
||||
Keep all other parts of the review unchanged.
|
||||
|
||||
Current Draft Review:
|
||||
```text
|
||||
{state['draft_review']}
|
||||
```
|
||||
"""
|
||||
response = await llm.ainvoke(HumanMessage(content=prompt))
|
||||
state['draft_review'] = response.content.strip()
|
||||
state['round'] += 1
|
||||
state["round"] += 1
|
||||
prompt = (
|
||||
f"Rewrite the part of the review that addresses the weakest criterion "
|
||||
f"('{state['weakest_criterion']}') to improve it. Keep the rest of the review unchanged.\n\n"
|
||||
f"Original Review:\n{state['draft_review']}\n\n"
|
||||
"Provide only the updated review."
|
||||
)
|
||||
response = await agent.ainvoke(
|
||||
{"messages": [HumanMessage(content=prompt)]},
|
||||
{"configurable": {"thread_id": "rewrite"}},
|
||||
)
|
||||
new_review = response["messages"][-1].content.strip()
|
||||
state["draft_review"] = new_review
|
||||
print("\n--- Rewritten Review (Round {}) ---".format(state["round"]))
|
||||
print(new_review)
|
||||
return state
|
||||
|
||||
# ---------- Graph ----------
|
||||
review_graph = StateGraph(CodeReviewState)
|
||||
review_graph.add_node("draft_review", draft_review)
|
||||
review_graph.add_node("reflect", reflect)
|
||||
review_graph.add_node("rewrite", rewrite)
|
||||
# ---------- Graph construction ----------
|
||||
builder = StateGraph(CodeReviewState)
|
||||
|
||||
review_graph.set_entry_point("draft_review")
|
||||
review_graph.add_edge("draft_review", "reflect")
|
||||
builder.add_node("draft_review", draft_review)
|
||||
builder.add_node("reflect", reflect)
|
||||
builder.add_node("rewrite", rewrite)
|
||||
|
||||
builder.add_edge(START, "draft_review")
|
||||
builder.add_edge("draft_review", "reflect")
|
||||
|
||||
# Conditional edges after reflect
|
||||
review_graph.add_conditional_edges(
|
||||
"reflect",
|
||||
lambda state: "END" if state["verdict"] == "ok" else "rewrite" if state["round"] < state["max_rounds"] else "END",
|
||||
)
|
||||
review_graph.add_edge("rewrite", "reflect")
|
||||
def reflect_conditional(state: CodeReviewState):
|
||||
if state["verdict"] == "ok":
|
||||
return END
|
||||
if state["round"] < state["max_rounds"]:
|
||||
return "rewrite"
|
||||
return END
|
||||
|
||||
compiled_graph = review_graph.compile()
|
||||
builder.add_conditional_edges("reflect", reflect_conditional)
|
||||
|
||||
# ---------- Tool ----------
|
||||
@tool
|
||||
async def run_code_review(code: str) -> Dict:
|
||||
"""Run a code review on the provided Python function."""
|
||||
builder.add_edge("rewrite", "reflect")
|
||||
|
||||
graph = builder.compile()
|
||||
|
||||
# ---------- Demo ----------
|
||||
async def main():
|
||||
# Sample function to review
|
||||
sample_code = """
|
||||
def sort_numbers(arr):
|
||||
return sorted(arr)
|
||||
"""
|
||||
initial_state: CodeReviewState = {
|
||||
"code": code,
|
||||
"code": sample_code.strip(),
|
||||
"draft_review": "",
|
||||
"criteria_scores": {},
|
||||
"weakest_criterion": "",
|
||||
@@ -147,37 +164,13 @@ async def run_code_review(code: str) -> Dict:
|
||||
"round": 0,
|
||||
"max_rounds": 2,
|
||||
}
|
||||
final_state = compiled_graph.invoke(initial_state)
|
||||
return {
|
||||
"final_review": final_state["draft_review"],
|
||||
"scores": final_state["criteria_scores"],
|
||||
"weakest_criterion": final_state["weakest_criterion"],
|
||||
"verdict": final_state["verdict"],
|
||||
"rounds": final_state["round"],
|
||||
}
|
||||
|
||||
# ---------- Agent ----------
|
||||
agent = create_deep_agent(
|
||||
model=llm,
|
||||
tools=[run_code_review],
|
||||
backend=backend,
|
||||
system_prompt="You are a helpful code review agent.",
|
||||
)
|
||||
|
||||
# ---------- Demo ----------
|
||||
async def main():
|
||||
# Sample function to review
|
||||
code = """
|
||||
def sort_numbers(arr):
|
||||
return sorted(arr)
|
||||
"""
|
||||
result = await run_code_review(code)
|
||||
print("\n=== Final Review ===\n")
|
||||
print(result["final_review"])
|
||||
print("\n=== Scores ===\n")
|
||||
for k, v in result["scores"].items():
|
||||
print(f"{k}: {v}")
|
||||
print(f"\nVerdict: {result['verdict']} (Rounds: {result['rounds']})")
|
||||
final_state = await graph.ainvoke(initial_state)
|
||||
print("\n=== Final State ===")
|
||||
print(f"Verdict: {final_state['verdict']}")
|
||||
print(f"Rounds performed: {final_state['round']}")
|
||||
print("\nFinal Review:")
|
||||
print(final_state["draft_review"])
|
||||
|
||||
if __name__ == "__main__":
|
||||
asyncio.run(main())
|
||||
Reference in New Issue
Block a user