README.md

metadata

language:
  - en

This repository contains the index files required for the SearchWikipediaTool of the bot-with-plan project.

It is based on the krasserm/wikipedia-2023-11-en-embed-mxbai-int8-binary dataset and contains the following files:

• faiss-ubinary.index: Faiss index file containing the binary embeddings
• usearch-int8-index: usearch index files containing the int8 embeddings
• document-url-mappings.sqlite: SQLite database file containing mappings from document URLs to text chunk indices

The following code snippet demonstrates how to use the index files with the SearchWikipediaTool:

from sentence_transformers import CrossEncoder, SentenceTransformer
from gba.client import Llama3Instruct, LlamaCppClient
from gba.tools import SearchWikipediaTool
from gba.utils import Scratchpad

llm_model = Llama3Instruct(llm=LlamaCppClient(url="http://localhost:8084/completion", temperature=-1))
embedding_model = SentenceTransformer("mixedbread-ai/mxbai-embed-large-v1", device="cuda")
rerank_model = CrossEncoder("mixedbread-ai/mxbai-rerank-large-v1", device="cuda")

search_wikipedia = SearchWikipediaTool(
   llm=llm_model,
   embedding_model=embedding_model,
   rerank_model=rerank_model,
   top_k_nodes=10,
   top_k_related_documents=1,
   top_k_related_nodes=3,
)

response = search_wikipedia.run(
   task="Search Wikipedia for the launch date of the first iPhone.",
   request="",
   scratchpad=Scratchpad(),
)