Awesome
🦜️🌐 WebLangChain
This repo is an example of performing retrieval using the entire internet as a document store.
Try it live: weblangchain.vercel.app
✅ Running locally
By default, WebLangChain uses Tavily to fetch content from webpages. You can get an API key from by signing up.
If you'd like to add or swap in different base retrievers (e.g. if you want to use your own data source), you can update the get_retriever()
method in main.py
.
- Install backend dependencies:
poetry install
. - Make sure to set your environment variables to configure the application:
export OPENAI_API_KEY=
export TAVILY_API_KEY=
# for Anthropic
# remove models from code if unused
ANTHROPIC_API_KEY=
# if you'd like to use the You.com retriever
export YDC_API_KEY=
# if you'd like to use the Google retriever
export GOOGLE_CSE_ID=
export GOOGLE_API_KEY=
# if you'd like to use the Kay.ai retriever
export KAY_API_KEY=
# for tracing
export LANGCHAIN_TRACING_V2=true
export LANGCHAIN_ENDPOINT="https://api.smith.langchain.com"
export LANGCHAIN_API_KEY=
export LANGCHAIN_PROJECT=
You will also need to put your Google Cloud credentials in a JSON file under .google_vertex_ai_credentials.json
in the main directory if you would like to
use Google Vertex as an option. If you're not using Vertex, you'll need to remove ChatVertexAI
from main.py
.
- Start the Python backend with
poetry run make start
. - Install frontend dependencies by running
cd nextjs
, thenyarn
. - Run the frontend with
yarn dev
for frontend. - Open localhost:3000 in your browser.
Under the hood, the chain is converted to a FastAPI server with various endpoints via LangServe.
This also includes a playground that you can use to interactively swap and configure various pieces of the chain.
You can find it running at http://localhost:8080/chat/playground
.
☕ Running locally (JS backend)
Note that LangServe is not currently supported in JS, and customization of the retriever and model, as well as the playground, are unavailable.
- Install frontend dependencies by running
cd nextjs
, thenyarn
. - Populate a
nextjs/.env.local
file with your own versions of keys from thenextjs/.env.example
file, and setNEXT_PUBLIC_API_BASE_URL
to"http://localhost:3000/api"
. - Run the app with
yarn dev
. - Open localhost:3000 in your browser.
⚙️ How it works
The general retrieval flow looks like this:
- Pull in raw content related to the user's initial query using a retriever that wraps Tavily's Search API.
- For subsequent conversation turns, we also rephrase the original query into a "standalone query" free of references to previous chat history.
- Because the size of the raw documents usually exceed the maximum context window size of the model, we perform additional contextual compression steps to filter what we pass to the model.
- First, we split retrieved documents using a text splitter.
- Then we use an embeddings filter to remove any chunks that do not meet a similarity threshold with the initial query.
- The retrieved context, the chat history, and the original question are passed to the LLM as context for the final generation.
Here's a LangSmith trace illustrating the above:
https://smith.langchain.com/public/f4493d9c-218b-404a-a890-31c15c56fff3/r
It's built using:
- Tavily as a retriever
- LangChain for orchestration
- LangServe to directly expose LangChain runnables as endpoints
- FastAPI
- Next.js for the frontend
🚀 Deployment
The live version is hosted on Fly.dev and Vercel.
The backend Python logic is found in main.py
, and the frontend Next.js app is under nextjs/
.