Let's create some demo corpus - Llama Index

In this notebook, we are going to show how to use BGE-M3 with LlamaIndex.

BGE-M3 is a hybrid multilingual retrieval model that supports over 100 languages and can handle input lengths of up to 8,192 tokens. The model can perform (i) dense retrieval, (ii) sparse retrieval, and (iii) multi-vector retrieval.

Getting Started

python

%pip install llama-index-indices-managed-bge-m3

python

%pip install llama-index

Creating BGEM3Index

python

from llama_index.core import Settings
from llama_index.core import Document
from llama_index.indices.managed.bge_m3 import BGEM3Index

Settings.chunk_size = 8192

python

# Let's create some demo corpus
sentences = [
    "BGE M3 is an embedding model supporting dense retrieval, lexical matching and multi-vector interaction.",
    "BM25 is a bag-of-words retrieval function that ranks a set of documents based on the query terms appearing in each document",
]
documents = [Document(doc_id=i, text=s) for i, s in enumerate(sentences)]

python

# Indexing with BGE-M3 model
index = BGEM3Index.from_documents(
    documents,
    weights_for_different_modes=[
        0.4,
        0.2,
        0.4,
    ],  # [dense_weight, sparse_weight, multi_vector_weight]
)

Retrieve relevant documents

python

retriever = index.as_retriever()
response = retriever.retrieve("What is BGE-M3?")

RAG with BGE-M3

python

query_engine = index.as_query_engine()
response = query_engine.query("What is BGE-M3?")