111 lines
3.5 KiB
Python
111 lines
3.5 KiB
Python
from typing import TypedDict, Dict
|
|
import inspect
|
|
|
|
from langgraph.graph import StateGraph, END
|
|
from langchain_ollama import Ollama
|
|
from langchain_core.prompts import ChatPromptTemplate
|
|
from langchain_core.output_parsers import JsonOutputParser
|
|
|
|
# Define the state
|
|
class CodeReviewState(TypedDict):
|
|
code: str
|
|
draft_review: str
|
|
criteria_scores: Dict[str, int]
|
|
weakest_criterion: str
|
|
verdict: str # "ok" | "needs_revision"
|
|
round: int
|
|
max_rounds: int
|
|
|
|
# LLM instance (Ollama)
|
|
llm = Ollama(model="llama3.1")
|
|
|
|
# Node: draft_review
|
|
|
|
def draft_review(state: CodeReviewState) -> Dict[str, str]:
|
|
prompt = ChatPromptTemplate.from_messages([
|
|
("system", "You are a senior Python developer. Write a concise code review for the given function. Provide 3-6 actionable points."),
|
|
("user", "Here is the function:\n{code}")
|
|
])
|
|
chain = prompt | llm
|
|
review = chain.invoke({"code": state["code"]})
|
|
return {"draft_review": review}
|
|
|
|
# Node: reflect
|
|
|
|
def reflect(state: CodeReviewState) -> Dict[str, object]:
|
|
prompt = ChatPromptTemplate.from_messages([
|
|
("system", """You are a code quality critic. Score the following review on four criteria: PEP8, type hints, edge cases, naming. Return a JSON with integer scores 0-10, the weakest criterion, and verdict \"ok\" or \"needs_revision\".\n""") ,
|
|
("user", "Review:\n{draft_review}")
|
|
])
|
|
parser = JsonOutputParser()
|
|
chain = prompt | llm | parser
|
|
result = chain.invoke({"draft_review": state["draft_review"]})
|
|
# result is a dict
|
|
return {
|
|
"criteria_scores": {
|
|
"pep8": result["pep8"],
|
|
"type_hints": result["type_hints"],
|
|
"edge_cases": result["edge_cases"],
|
|
"naming": result["naming"],
|
|
},
|
|
"weakest_criterion": result["weakest_criterion"],
|
|
"verdict": result["verdict"],
|
|
}
|
|
|
|
# Node: rewrite
|
|
|
|
def rewrite(state: CodeReviewState) -> Dict[str, str]:
|
|
# Increment round
|
|
state["round"] += 1
|
|
prompt = ChatPromptTemplate.from_messages([
|
|
("system", "You are a senior Python developer. Rewrite the review to improve the section about {weakest_criterion}. Keep other points unchanged."),
|
|
("user", "Original review:\n{draft_review}")
|
|
])
|
|
chain = prompt | llm
|
|
new_review = chain.invoke({"weakest_criterion": state["weakest_criterion"], "draft_review": state["draft_review"]})
|
|
return {"draft_review": new_review}
|
|
|
|
# Build the graph
|
|
builder = StateGraph(CodeReviewState)
|
|
builder.add_node("draft_review", draft_review)
|
|
builder.add_node("reflect", reflect)
|
|
builder.add_node("rewrite", rewrite)
|
|
|
|
builder.add_edge("draft_review", "reflect")
|
|
# Conditional edge after reflect
|
|
builder.add_conditional_edges(
|
|
"reflect",
|
|
lambda state: "END" if state["verdict"] == "ok" else "rewrite",
|
|
)
|
|
builder.add_edge("rewrite", "reflect")
|
|
|
|
builder.set_entry_point("draft_review")
|
|
builder.set_finish_point("END")
|
|
|
|
graph = builder.compile()
|
|
|
|
# Demo
|
|
if __name__ == "__main__":
|
|
def sort_numbers(arr):
|
|
return sorted(arr)
|
|
|
|
code = inspect.getsource(sort_numbers)
|
|
initial_state: CodeReviewState = {
|
|
"code": code,
|
|
"draft_review": "",
|
|
"criteria_scores": {},
|
|
"weakest_criterion": "",
|
|
"verdict": "",
|
|
"round": 0,
|
|
"max_rounds": 2,
|
|
}
|
|
result = graph.invoke(initial_state)
|
|
print("\n--- Draft Review ---")
|
|
print(result["draft_review"])
|
|
print("\n--- Scores ---")
|
|
print(result["criteria_scores"])
|
|
print("\n--- Verdict ---")
|
|
print(result["verdict"])
|
|
print("\n--- Round ---")
|
|
print(result["round"])
|