Skip to main content


This template demonstrates the multi-vector indexing strategy proposed by Chen, et. al.'s Dense X Retrieval: What Retrieval Granularity Should We Use?. The prompt, which you can try out on the hub, directs an LLM to generate de-contextualized "propositions" which can be vectorized to increase the retrieval accuracy. You can see the full definition in


For this demo, we index a simple academic paper using the RecursiveUrlLoader, and store all retriever information locally (using chroma and a bytestore stored on the local filesystem). You can modify the storage layer in

Environment Setup​

Set the OPENAI_API_KEY environment variable to access gpt-3.5 and the OpenAI Embeddings classes.


Create the index by running the following:

poetry install
poetry run python propositional_retrieval/


To use this package, you should first have the LangChain CLI installed:

pip install -U langchain-cli

To create a new LangChain project and install this as the only package, you can do:

langchain app new my-app --package propositional-retrieval

If you want to add this to an existing project, you can just run:

langchain app add propositional-retrieval

And add the following code to your file:

from propositional_retrieval import chain

add_routes(app, chain, path="/propositional-retrieval")

(Optional) Let's now configure LangSmith. LangSmith will help us trace, monitor and debug LangChain applications. You can sign up for LangSmith here. If you don't have access, you can skip this section

export LANGCHAIN_API_KEY=<your-api-key>
export LANGCHAIN_PROJECT=<your-project> # if not specified, defaults to "default"

If you are inside this directory, then you can spin up a LangServe instance directly by:

langchain serve

This will start the FastAPI app with a server is running locally at http://localhost:8000

We can see all templates at We can access the playground at

We can access the template from code with:

from langserve.client import RemoteRunnable

runnable = RemoteRunnable("http://localhost:8000/propositional-retrieval")

Help us out by providing feedback on this documentation page: