llama-index-packs-raptor 0.1.1

llama-index packs raptor integration

Raptor Retriever LlamaPack

This LlamaPack shows how to use an implementation of RAPTOR with llama-index, leveraging the RAPTOR pack.

RAPTOR works by recursively clustering and summarizing clusters in layers for retrieval.

There two retrieval modes:

• tree_traversal -- traversing the tree of clusters, performing top-k at each level in the tree.
• collapsed -- treat the entire tree as a giant pile of nodes, perform simple top-k.

See the paper for full algorithm details.

CLI Usage

You can download llamapacks directly using llamaindex-cli, which comes installed with the llama-index python package:

llamaindex-cli download-llamapack RaptorPack --download-dir ./raptor_pack

You can then inspect/modify the files at ./raptor_pack and use them as a template for your own project.

Code Usage

You can alternaitvely install the package:

pip install llama-index-packs-raptor

Then, you can import and initialize the pack! This will perform clustering and summarization over your data.

from llama_index.packs.raptor import RaptorPack

pack = RaptorPack(documents, llm=llm, embed_model=embed_model)

The run() function is a light wrapper around retriever.retrieve().

nodes = pack.run(
    "query",
    mode="collapsed",  # or tree_traversal
)

You can also use modules individually.

# get the retriever
retriever = pack.retriever

Persistence

The RaptorPack comes with the RaptorRetriever, which offers ways of saving/reloading!

If you are using a remote vector-db, just pass it in

# Pack usage
pack = RaptorPack(..., vector_store=vector_store)

# RaptorRetriever usage
retriever = RaptorRetriever(..., vector_store=vector_store)

Then, to re-connect, just pass in the vector store again and an empty list of documents

# Pack usage
pack = RaptorPack([], ..., vector_store=vector_store)

# RaptorRetriever usage
retriever = RaptorRetriever([], ..., vector_store=vector_store)

Check out the notebook here for complete details!.