search

Search tools

This package contains tools for internet and Wikipedia searches.

Search internet tool

The SearchInternetTool is a RAG-based search tool that performs internet searches across multiple search engines and synthesizes concise responses using a large language model (LLM).

The tool does not require API keys as it utilizes a local SearXNG meta search instance to query various search engines. Results from the search engines are retrieved from the internet, filtered and re-ranked by relevance and finally passed to the LLM to generate a concise response.

For a more detailed explanation of the implementation details refer to the section search internet tool implementation.

To use the SearchInternetTool you need to set up a local SearXNG instance and provide the endpoint URL to the tool.

Setup

Setup a local SearXNG instance using the official docker container (see also SearXNG docs).

docker run \
  --name searxng \
  -d -p 8080:8080 \
  -v "${PWD}/.searxng:/etc/searxng" \
  -e "BASE_URL=http://localhost:8080" \
  -e "INSTANCE_NAME=my-instance" \
  searxng/searxng:2024.5.24-75e4b6512

The .searxng/settings.yaml file used in this project has been modified to additionally support json mode:

   search:
     formats:
     - html
     - json

See getting started for instructions how to serve the Llama 3 model used by these tools.

Usage

from gba.tools.search import create_search_internet_tool

search_internet = create_search_internet_tool(
    llama3_endpoint="http://localhost:8084/completion",
    searxng_endpoint="http://localhost:8080",
)

response = search_internet.search(
    query="When was the video game 'The Last of Us' released"
)

Parameters

• top_k_documents: number of top-k webpages to select from the search results (default: 3)
• top_k_nodes_per_document: number of top-k relevant text nodes to select from each webpage for generating the response (default: 5)
• top_k_snippets: number of top-k webpage snippets to include (default: top_k_documents)

The complete list of parameters can be found in the SearchInternetTool class.

Search Wikipedia tool

The SearchWikipediaTool is a RAG-based search tool designed for efficient search in a local Wikipedia dataset.

The tool utilizes multiple locally stored, quantized search indices for memory and runtime-efficient nearest neighbor searches in the dataset. Given a search query, the tool retrieves the most relevant text nodes from Wikipedia articles and synthesizes responses using a large language model (LLM).

Note: the dataset has a knowledge cutoff of November 2023.

For a more detailed explanation of the implementation details refer to the section search wikipedia tool implementation.

Usage

from gba.tools.search import create_search_wikipedia_tool

search_wikipedia = create_search_wikipedia_tool(
   llama3_endpoint="http://localhost:8084/completion",
)

response = search_wikipedia.search(
   search="Search Wikipedia for the launch date of the first iPhone."
)

Parameters

• top_k_nodes: number of top-k text nodes to select from the initial search results (default: 10)
• top_k_related_documents: number of top-k related documents to select from the initial search results (default: 1)
• top_k_related_nodes: number of top-k text nodes to select from the related documents (default: 3)

The complete list of parameters can be found in the SearchWikipediaTool class.