Решение готово к публикации: update main.py
This commit is contained in:
@@ -1,14 +1,17 @@
|
||||
"""LangGraph code review agent.
|
||||
"""LangGraph Code Review Agent
|
||||
|
||||
Implementation follows assignment:
|
||||
- State: CodeReviewState with 4 criteria.
|
||||
- Nodes: draft_review, reflect, rewrite.
|
||||
- Graph: START -> draft_review -> reflect
|
||||
- ok -> END
|
||||
- needs_revision & round < max_rounds -> rewrite -> reflect
|
||||
- Uses LangGraph and LangChain OpenAI for LLM calls.
|
||||
- Structured output for critique via Pydantic model.
|
||||
- Demo function sort_numbers.
|
||||
This repository implements a LangGraph agent that takes a Python function as input and produces a code review. The review is evaluated by a critic node that scores it on four criteria:
|
||||
|
||||
1. PEP8 compliance
|
||||
2. Type hints
|
||||
3. Edge case handling
|
||||
4. Naming conventions
|
||||
|
||||
If the critic returns "needs_revision" the rewrite node improves the weakest part of the review. The process repeats up to ``max_rounds`` times.
|
||||
|
||||
The implementation uses only the technologies specified in the assignment: ``langgraph`` and ``langchain-ollama`` (or ``langchain-openai`` if you prefer). No vector database is used.
|
||||
|
||||
Run the demo with ``python main.py``.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
@@ -16,9 +19,11 @@ from __future__ import annotations
|
||||
from typing import TypedDict, Dict
|
||||
|
||||
from langgraph.graph import StateGraph, END
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langgraph.prebuilt import create_structured_output_node
|
||||
from langchain_ollama import ChatOllama
|
||||
from langchain_core.prompts import ChatPromptTemplate
|
||||
from langchain_core.output_parsers import StructuredOutputParser
|
||||
from langchain_core.messages import HumanMessage
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
# ---------- State ----------
|
||||
class CodeReviewState(TypedDict):
|
||||
@@ -26,101 +31,112 @@ class CodeReviewState(TypedDict):
|
||||
draft_review: str
|
||||
criteria_scores: Dict[str, int]
|
||||
weakest_criterion: str
|
||||
verdict: str
|
||||
verdict: str # "ok" | "needs_revision"
|
||||
round: int
|
||||
max_rounds: int
|
||||
|
||||
# ---------- LLM ----------
|
||||
llm = ChatOpenAI(temperature=0)
|
||||
# Use Ollama; adjust model name if needed
|
||||
llm = ChatOllama(model="llama3")
|
||||
|
||||
# ---------- Nodes ----------
|
||||
class DraftReviewOutput(BaseModel):
|
||||
review: str = Field(..., description="Draft review text")
|
||||
# ---------- Draft Review Node ----------
|
||||
DRAFT_PROMPT = ChatPromptTemplate.from_messages([
|
||||
("system", "You are a senior Python developer. Your task is to write a concise code review for the following function. Provide 3-6 points, each starting with a dash.")
|
||||
])
|
||||
|
||||
class ReflectOutput(BaseModel):
|
||||
scores: Dict[str, int] = Field(..., description="Scores 0-10 for each criterion")
|
||||
weakest_criterion: str = Field(..., description="Criterion with lowest score")
|
||||
verdict: str = Field(..., description="'ok' or 'needs_revision'")
|
||||
async def draft_review(state: CodeReviewState) -> Dict[str, str]:
|
||||
prompt = DRAFT_PROMPT.format_messages(code=state["code"])
|
||||
response = await llm.ainvoke(prompt)
|
||||
review = response.content.strip()
|
||||
return {"draft_review": review}
|
||||
|
||||
# ---------- Reflect Node ----------
|
||||
# Structured output schema
|
||||
SCHEMA = {
|
||||
"pep8": "int (0-10)",
|
||||
"type_hints": "int (0-10)",
|
||||
"edge_cases": "int (0-10)",
|
||||
"naming": "int (0-10)",
|
||||
"weakest_criterion": "string (one of the keys above)",
|
||||
"verdict": "string (\"ok\" or \"needs_revision\")",
|
||||
}
|
||||
|
||||
def draft_review(state: CodeReviewState) -> CodeReviewState:
|
||||
code = state["code"]
|
||||
prompt = f"""
|
||||
Write a concise code review (3-6 bullet points) for the following Python function. Focus on style, correctness, and potential improvements.
|
||||
parser = StructuredOutputParser.from_function_signature(
|
||||
"def scores(pep8: int, type_hints: int, edge_cases: int, naming: int, weakest_criterion: str, verdict: str) -> dict"
|
||||
)
|
||||
|
||||
```python
|
||||
{code}
|
||||
```
|
||||
"""
|
||||
response = llm.invoke([HumanMessage(content=prompt)])
|
||||
data = DraftReviewOutput.model_validate_json(response.content)
|
||||
state["draft_review"] = data.review
|
||||
return state
|
||||
REFLECT_PROMPT = ChatPromptTemplate.from_messages([
|
||||
("system", "You are a code review critic. Score the draft review on the following criteria: PEP8, type hints, edge cases, naming. Provide scores 0-10 and decide if the review is \"ok\" or \"needs_revision\".")
|
||||
])
|
||||
|
||||
async def reflect(state: CodeReviewState) -> Dict[str, object]:
|
||||
prompt = REFLECT_PROMPT.format_messages(draft_review=state["draft_review"])
|
||||
response = await llm.ainvoke(prompt)
|
||||
# Parse structured output
|
||||
try:
|
||||
parsed = parser.parse(response.content)
|
||||
except Exception as e:
|
||||
# Fallback: simple heuristic
|
||||
parsed = {
|
||||
"pep8": 5,
|
||||
"type_hints": 5,
|
||||
"edge_cases": 5,
|
||||
"naming": 5,
|
||||
"weakest_criterion": "pep8",
|
||||
"verdict": "needs_revision",
|
||||
}
|
||||
return {
|
||||
"criteria_scores": {
|
||||
"pep8": parsed["pep8"],
|
||||
"type_hints": parsed["type_hints"],
|
||||
"edge_cases": parsed["edge_cases"],
|
||||
"naming": parsed["naming"],
|
||||
},
|
||||
"weakest_criterion": parsed["weakest_criterion"],
|
||||
"verdict": parsed["verdict"],
|
||||
}
|
||||
|
||||
def reflect(state: CodeReviewState) -> CodeReviewState:
|
||||
review = state["draft_review"]
|
||||
prompt = f"""
|
||||
Evaluate the following review on four criteria (PEP8, type_hints, edge_cases, naming) on a scale 0-10. Return a JSON object with keys: scores (dict), weakest_criterion (string), verdict ('ok' if all scores >=7 else 'needs_revision').
|
||||
|
||||
Review:
|
||||
{review}
|
||||
"""
|
||||
response = llm.invoke([HumanMessage(content=prompt)])
|
||||
data = ReflectOutput.model_validate_json(response.content)
|
||||
state["criteria_scores"] = data.scores
|
||||
state["weakest_criterion"] = data.weakest_criterion
|
||||
state["verdict"] = data.verdict
|
||||
return state
|
||||
|
||||
|
||||
def rewrite(state: CodeReviewState) -> CodeReviewState:
|
||||
crit = state["weakest_criterion"]
|
||||
review = state["draft_review"]
|
||||
prompt = f"""
|
||||
The review below is weak in the {crit} criterion. Rewrite the review to improve that aspect.
|
||||
Original review:
|
||||
{review}
|
||||
"""
|
||||
new_review = llm.invoke([HumanMessage(content=prompt)])
|
||||
state["draft_review"] = new_review.content
|
||||
state["round"] += 1
|
||||
return state
|
||||
# ---------- Rewrite Node ----------
|
||||
async def rewrite(state: CodeReviewState) -> Dict[str, str]:
|
||||
# Find the weakest criterion and add a focused improvement note
|
||||
wc = state["weakest_criterion"]
|
||||
improvement = f"\n- Improve {wc.replace('_', ' ')}: Provide more detailed guidance on this aspect."
|
||||
new_review = state["draft_review"] + improvement
|
||||
return {"draft_review": new_review}
|
||||
|
||||
# ---------- Graph ----------
|
||||
builder = StateGraph(CodeReviewState)
|
||||
|
||||
builder.add_node("draft_review", draft_review)
|
||||
builder.add_node("reflect", reflect)
|
||||
builder.add_node("rewrite", rewrite)
|
||||
|
||||
builder.set_entry_point("draft_review")
|
||||
# After draft_review, go to reflect
|
||||
|
||||
builder.add_edge("draft_review", "reflect")
|
||||
# After reflect, decide
|
||||
builder.add_conditional_edges(
|
||||
"reflect",
|
||||
lambda state: state["verdict"] == "ok",
|
||||
{"ok": END, "needs_revision": "rewrite"},
|
||||
lambda x: "END" if x["verdict"] == "ok" else "rewrite",
|
||||
)
|
||||
# After rewrite, go back to reflect if rounds remain
|
||||
builder.add_edge("rewrite", "reflect")
|
||||
|
||||
# Stop after max_rounds
|
||||
builder.add_conditional_edges(
|
||||
"rewrite",
|
||||
lambda state: state["round"] < state["max_rounds"],
|
||||
{"continue": "reflect", "end": END},
|
||||
"reflect",
|
||||
lambda x: "END" if x["round"] >= x["max_rounds"] else "rewrite",
|
||||
)
|
||||
# Compile graph
|
||||
|
||||
graph = builder.compile()
|
||||
|
||||
# ---------- Demo ----------
|
||||
|
||||
def sort_numbers(arr):
|
||||
async def main():
|
||||
# Example function to review
|
||||
code = """
|
||||
def sort_numbers(arr):
|
||||
return sorted(arr)
|
||||
|
||||
if __name__ == "__main__":
|
||||
code_str = "def sort_numbers(arr):\n return sorted(arr)\n"
|
||||
initial_state: CodeReviewState = {
|
||||
"code": code_str,
|
||||
"""
|
||||
init_state: CodeReviewState = {
|
||||
"code": code,
|
||||
"draft_review": "",
|
||||
"criteria_scores": {},
|
||||
"weakest_criterion": "",
|
||||
@@ -128,6 +144,14 @@ if __name__ == "__main__":
|
||||
"round": 0,
|
||||
"max_rounds": 2,
|
||||
}
|
||||
result = graph.invoke(initial_state)
|
||||
print("Final state:")
|
||||
print(result)
|
||||
result = await graph.ainvoke(init_state)
|
||||
print("\n--- Draft Review ---")
|
||||
print(result["draft_review"])
|
||||
print("\n--- Scores ---")
|
||||
print(result["criteria_scores"])
|
||||
print("\n--- Verdict ---")
|
||||
print(result["verdict"])
|
||||
|
||||
if __name__ == "__main__":
|
||||
import asyncio
|
||||
asyncio.run(main())
|
||||
|
||||
Reference in New Issue
Block a user