136 lines
4.5 KiB
Python
136 lines
4.5 KiB
Python
import os
|
||
import asyncio
|
||
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
|
||
from langchain_chroma import Chroma
|
||
from langchain_core.documents import Document
|
||
from langchain_text_splitters import RecursiveCharacterTextSplitter
|
||
from langchain.tools import tool
|
||
from deepagents import create_deep_agent
|
||
from deepagents.backends import FilesystemBackend, LocalShellBackend, CompositeBackend
|
||
from langchain_core.messages import HumanMessage
|
||
from tavily import TavilySearchResults
|
||
|
||
# ---------------------
|
||
# Configuration
|
||
# ---------------------
|
||
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
|
||
TAVILY_API_KEY = os.getenv("TAVILY_API_KEY")
|
||
|
||
# ---------------------
|
||
# Vector store utilities
|
||
# ---------------------
|
||
|
||
def create_vectorstore(persist_directory: str = "./chroma_db") -> Chroma:
|
||
"""Create or load a Chroma vector store with OpenAI embeddings."""
|
||
embeddings = OpenAIEmbeddings(
|
||
model="text-embedding-3-small",
|
||
base_url="https://openrouter.ai/api/v1",
|
||
api_key=OPENAI_API_KEY,
|
||
)
|
||
return Chroma(
|
||
collection_name="knowledge",
|
||
embedding_function=embeddings,
|
||
persist_directory=persist_directory,
|
||
)
|
||
|
||
|
||
def load_documents(directory: str, vectorstore: Chroma) -> None:
|
||
"""Load .txt and .md files from *directory* into *vectorstore* using chunking."""
|
||
splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
|
||
docs = []
|
||
for root, _, files in os.walk(directory):
|
||
for fname in files:
|
||
if fname.lower().endswith(('.txt', '.md')):
|
||
path = os.path.join(root, fname)
|
||
with open(path, "r", encoding="utf-8") as f:
|
||
text = f.read()
|
||
# Create Document objects
|
||
docs.extend(
|
||
[Document(page_content=chunk, metadata={"source": path}) for chunk in splitter.split_text(text)]
|
||
)
|
||
if docs:
|
||
vectorstore.add_documents(docs)
|
||
vectorstore.persist()
|
||
|
||
# ---------------------
|
||
# Tools
|
||
# ---------------------
|
||
|
||
vectorstore = create_vectorstore()
|
||
|
||
# Ensure we have some data loaded – load from ./documents if collection empty
|
||
if len(vectorstore.get_all_documents()) == 0:
|
||
load_documents("./documents", vectorstore)
|
||
|
||
@tool
|
||
def search_local_kb(query: str, top_k: int = 3) -> str:
|
||
"""Search the local knowledge base for relevant passages."""
|
||
docs = vectorstore.similarity_search(query, k=top_k)
|
||
if not docs:
|
||
return "[Local KB] No relevant information found."
|
||
result = "\n\n".join([f"{d.metadata.get('source', 'unknown')}\n{d.page_content}" for d in docs])
|
||
return f"[Local KB]\n{result}"
|
||
|
||
@tool
|
||
def web_search(query: str) -> str:
|
||
"""Perform a web search using Tavily and return top results."""
|
||
results = TavilySearchResults(query=query, max_results=3, api_key=TAVILY_API_KEY)
|
||
if not results.results:
|
||
return "[Web Search] No results found."
|
||
snippets = []
|
||
for r in results.results:
|
||
snippets.append(f"{r.get('title', 'No title')}\n{r.get('content', 'No content')}\nURL: {r.get('url', '')}")
|
||
return f"[Web Search]\n\n".join(snippets)
|
||
|
||
# ---------------------
|
||
# Agent setup
|
||
# ---------------------
|
||
|
||
llm = ChatOpenAI(
|
||
model="openai/gpt-oss-20b:free",
|
||
base_url="https://openrouter.ai/api/v1",
|
||
api_key=OPENAI_API_KEY,
|
||
temperature=0.0,
|
||
)
|
||
|
||
backend = CompositeBackend([
|
||
LocalShellBackend(workspace_dir="./workspace"),
|
||
FilesystemBackend(),
|
||
])
|
||
|
||
system_prompt = (
|
||
"You are a helpful RAG agent. For questions about local documents, use the tool `search_local_kb`. "
|
||
"For current news or facts that may not be in your local knowledge base, use `web_search`. "
|
||
"Return the answer prefixed with either `[Local KB]` or `[Web Search]` to indicate the source."
|
||
)
|
||
|
||
agent = create_deep_agent(
|
||
model=llm,
|
||
tools=[search_local_kb, web_search],
|
||
backend=backend,
|
||
system_prompt=system_prompt,
|
||
)
|
||
|
||
# ---------------------
|
||
# CLI loop
|
||
# ---------------------
|
||
|
||
async def main():
|
||
print("RAG Agent ready. Type your question (or 'exit' to quit).")
|
||
while True:
|
||
user_input = input("\nQuery: ")
|
||
if user_input.lower() in {"exit", "quit", "q"}:
|
||
print("Goodbye!")
|
||
break
|
||
# Invoke agent
|
||
result = await agent.ainvoke(
|
||
{"messages": [HumanMessage(content=user_input)]},
|
||
{"configurable": {"thread_id": "session-1"}},
|
||
)
|
||
# The agent returns a list of messages; get the last one
|
||
reply = result["messages"][-1].content
|
||
print(f"\n{reply}")
|
||
|
||
if __name__ == "__main__":
|
||
asyncio.run(main())
|