🦜️🌐 WebLangChain

Using QianWen LLM via Tavily API to fetch webpage content and pass it to the LLM for generation.

✅ HuggingFaceEmbeddings

Load local models with HuggingFaceEmbeddings to save on invocation costs and circumvent network access restrictions.

✅ Custom Retriever Demo

YouRetriever demonstrates an example of retrieving information from multiple web pages, storing it in a vector database, and serving as a retriever.

✅ 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.

1. Install backend dependencies: poetry install.
2. 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=

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.

⚙️ How it works

The general retrieval flow looks like this:

1. 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.
2. 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.
3. 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
• Chroma as embedding database