Перемотка и контрольные точки: как создать исследовательского чат-агента с LangGraph

!pip -q install -U langgraph langchain langchain-google-genai google-generativeai typing_extensions
!pip -q install "requests==2.32.4"

import os
import json
import textwrap
import getpass
import time
from typing import Annotated, List, Dict, Any, Optional


from typing_extensions import TypedDict


from langchain.chat_models import init_chat_model
from langchain_core.messages import BaseMessage
from langchain_core.tools import tool


from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.prebuilt import ToolNode, tools_condition


import requests
from requests.adapters import HTTPAdapter, Retry


if not os.environ.get("GOOGLE_API_KEY"):
   os.environ["GOOGLE_API_KEY"] = getpass.getpass(" Enter your Google API Key (Gemini): ")


llm = init_chat_model("google_genai:gemini-2.0-flash")

WIKI_SEARCH_URL = "https://en.wikipedia.org/w/api.php"


_session = requests.Session()
_session.headers.update({
   "User-Agent": "LangGraph-Colab-Demo/1.0 (contact: example@example.com)",
   "Accept": "application/json",
})
retry = Retry(
   total=5, connect=5, read=5, backoff_factor=0.5,
   status_forcelist=(429, 500, 502, 503, 504),
   allowed_methods=("GET", "POST")
)
_session.mount("https://", HTTPAdapter(max_retries=retry))
_session.mount("http://", HTTPAdapter(max_retries=retry))


def _wiki_search_raw(query: str, limit: int = 3) -> List[Dict[str, str]]:
   """
   Use MediaWiki search API with:
     - origin='*' (good practice for CORS)
     - Polite UA + retries
   Returns compact list of {title, snippet_html, url}.
   """
   params = {
       "action": "query",
       "list": "search",
       "format": "json",
       "srsearch": query,
       "srlimit": limit,
       "srprop": "snippet",
       "utf8": 1,
       "origin": "*",
   }
   r = _session.get(WIKI_SEARCH_URL, params=params, timeout=15)
   r.raise_for_status()
   data = r.json()
   out = []
   for item in data.get("query", {}).get("search", []):
       title = item.get("title", "")
       page_url = f"https://en.wikipedia.org/wiki/{title.replace(' ', '_')}"
       snippet = item.get("snippet", "")
       out.append({"title": title, "snippet_html": snippet, "url": page_url})
   return out


@tool
def wiki_search(query: str) -> List[Dict[str, str]]:
   """Search Wikipedia and return up to 3 results with title, snippet_html, and url."""
   try:
       results = _wiki_search_raw(query, limit=3)
       return results if results else [{"title": "No results", "snippet_html": "", "url": ""}]
   except Exception as e:
       return [{"title": "Error", "snippet_html": str(e), "url": ""}]


TOOLS = [wiki_search]

class State(TypedDict):
   messages: Annotated[list, add_messages]


graph_builder = StateGraph(State)


llm_with_tools = llm.bind_tools(TOOLS)


SYSTEM_INSTRUCTIONS = textwrap.dedent("""
You are ResearchBuddy, a careful research assistant.
- If the user asks you to "research", "find info", "latest", "web", or references a library/framework/product,
 you SHOULD call the `wiki_search` tool at least once before finalizing your answer.
- When you call tools, be concise in the text you produce around the call.
- After receiving tool results, cite at least the page titles you used in your summary.
""").strip()


def chatbot(state: State) -> Dict[str, Any]:
   """Single step: call the LLM (with tools bound) on the current messages."""
   return {"messages": [llm_with_tools.invoke(state["msgs"])]]}


graph_builder.add_node("chatbot", chatbot)


memory = InMemorySaver()
graph = graph_builder.compile(checkpointer=memory)

def print_last_message(event: Dict[str, Any]):
   """Pretty-print the last message in an event if available."""
   if "messages" in event and event["messages"]:
       msg = event["messages"][-1]
       try:
           if isinstance(msg, BaseMessage):
               msg.pretty_print()
           else:
               role = msg.get("role", "unknown")
               content = msg.get("content", "")
               print(f"\n[{role.upper()}]\n{content}\n")
       except Exception:
           print(str(msg))


def show_state_history(cfg: Dict[str, Any]) -> List[Any]:
   """Print a concise view of checkpoints; return the list as well."""
   history = list(graph.get_state_history(cfg))
   print("\n===  State history (most recent first) ===")
   for i, st in enumerate(history):
       n = st.next
       n_txt = f"{n}" if n else "()"
       print(f"{i:02d}) NumMessages={len(st.values.get('messages', []))}  Next={n_txt}")
   print("=== End history ===\n")
   return history


def pick_checkpoint_by_next(history: List[Any], node_name: str = "tools") -> Optional[Any]:
   """Pick the first checkpoint whose `next` includes a given node (e.g., 'tools')."""
   for st in history:
       nxt = tuple(st.next) if st.next else tuple()
       if node_name in nxt:
           return st
   return None

config = {"configurable": {"thread_id": "demo-thread-1"}}


first_turn = {
   "messages": [
       {"role": "system", "content": SYSTEM_INSTRUCTIONS},
       {"role": "user", "content": "I'm learning LangGraph. Could you do some research on it for me?"},
   ]
}


print("\n====================  STEP 1: First user turn ====================")
events = graph.stream(first_turn, config, stream_mode="values")
for ev in events:
   print_last_message(ev)


second_turn = {
   "messages": [
       {"role": "user", "content": "Ya. Maybe I'll build an agent with it!"}
   ]
}


print("\n====================  STEP 2: Second user turn ====================")
events = graph.stream(second_turn, config, stream_mode="values")
for ev in events:
   print_last_message(ev)

print("\n====================  REPLAY: Full state history ====================")
history = show_state_history(config)


to_replay = pick_checkpoint_by_next(history, node_name="tools")
if to_replay is None:
   to_replay = history[min(2, len(history) - 1)]


print("Chosen checkpoint to resume from:")
print("  Next:", to_replay.next)
print("  Config:", to_replay.config)


print("\n====================  RESUME from chosen checkpoint ====================")
for ev in graph.stream(None, to_replay.config, stream_mode="vals"):
   print_last_message(ev)


MANUAL_INDEX = None 
if MANUAL_INDEX is not None and 0 <= MANUAL_INDEX < len(history):
   chosen = history[MANUAL_INDEX]
   print(f"\n====================  MANUAL RESUME @ index {MANUAL_INDEX} ====================")
   print("Next:", chosen.next)
   print("Config:", chosen.config)
   for ev in graph.stream(None, chosen.config, stream_mode="values"):
       print_last_message(ev)


print("\n Done. You added steps, replayed history, and resumed from a prior checkpoint.")