Files
povtornyy-ekzamen-2-sravnit…/src/nodes.py
T

99 lines
3.6 KiB
Python

import os
from typing import Dict, List
from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from langchain_tavily import TavilySearchTool
from langgraph.prebuilt import create_chat_agent
from langgraph import add_messages, StateGraph
from src.state import CompareState
load_dotenv()
# LLM and Tavily tool
llm = ChatOpenAI(temperature=0.7)
tavily = TavilySearchTool(api_key=os.getenv("TAVILY_API_KEY"))
# Helper to format findings for LLM
def format_findings(findings: Dict[str, List[str]]) -> str:
parts = []
for entity, notes in findings.items():
parts.append(f"**{entity}**:")
for note in notes:
parts.append(f"- {note}")
return "\n".join(parts)
# Node: Generate comparison criteria
def plan_criteria(state: CompareState) -> CompareState:
entities = state["entities"]
prompt = (
f"Generate 3-5 concise comparison criteria for the following entities: "
f"{', '.join(entities)}. Return a numbered list."
)
response = llm.invoke(prompt)
# Parse numbered list
criteria = []
for line in response.splitlines():
line = line.strip()
if line:
# Remove leading numbers
if line[0].isdigit() and (len(line) > 1 and line[1] in ". "):
line = line[2:].strip()
criteria.append(line)
state["criteria"] = criteria
# Initialize findings dict
state["findings"] = {entity: [] for entity in entities}
return state
# Node: Research one entity-criterion pair
def research_entity(state: CompareState) -> CompareState:
entities = state["entities"]
criteria = state["criteria"]
findings = state["findings"]
# Find next entity needing research
for entity in entities:
if len(findings[entity]) < len(criteria):
idx = len(findings[entity])
criterion = criteria[idx]
query = f"{entity} {criterion}"
# Tavily search
results = tavily.invoke({"query": query, "max_results": 3})
# Take first result snippet
if results and "results" in results and len(results["results"]) > 0:
snippet = results["results"][0]["snippet"]
source = results["results"][0]["url"]
note = f"{criterion}: {snippet} (Source: {source})"
else:
note = f"{criterion}: No recent information found."
findings[entity].append(note)
break
state["findings"] = findings
return state
# Node: Build cohesive research brief
def build_brief(state: CompareState) -> CompareState:
findings_text = format_findings(state["findings"])
prompt = (
f"Using the following findings, write a cohesive research brief that summarizes "
f"the strengths and weaknesses of each entity. The brief should be clear, "
f"structured, and suitable for a technical audience.\n\n"
f"Findings:\n{findings_text}"
)
brief = llm.invoke(prompt)
state["final_brief"] = brief
return state
# Node: Generate verdict/recommendation
def verdict(state: CompareState) -> CompareState:
brief = state["final_brief"]
criteria = state["criteria"]
prompt = (
f"Based on the research brief below and the comparison criteria, provide a "
f"clear recommendation on which entity is best suited for a typical use case. "
f"Explain your reasoning in 2-4 sentences.\n\n"
f"Research Brief:\n{brief}\n\n"
f"Criteria:\n- " + "\n- ".join(criteria)
)
recommendation = llm.invoke(prompt)
state["verdict"] = recommendation
return state