Raptor Retriever LlamaPack

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!.

Configure Summary Module

Using the SummaryModule you can configure how the Raptor Pack does summaries and how many workers are applied to summaries.

You can configure the LLM.

You can configure summary_prompt. This will change the prompt sent to your LLM to summarize you docs.

You can configure num_workers, which will influence the number of workers or rather async semaphores allowing more summaries to process simulatneously.
This might affect openai or other LLm provider API limits, be aware.

from llama_index.packs.raptor.base import SummaryModule
from llama_index.packs.raptor import RaptorRetriever

summary_prompt = "As a professional summarizer, create a concise and comprehensive summary of the provided text, be it an article, post, conversation, or passage with as much detail as possible."

### Adding SummaryModule you can configure the summary prompt and number of workers doing summaries.
summary_module = SummaryModule(
    llm=llm, summary_prompt=summary_prompt, num_workers=16
)

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