SearchAugmentedLLM empowers Large Language Models (LLMs) information from the web.
Given a user query, it performs a Google search, processes the top search results, chunks the content, ranks by relevance, and returns the most pertinent text to provide context for improved LLM responses. This tool is ideal for Retrieval Augmented Generation (RAG) applications.

Beta Warning

This project is in beta, and may not have the quality you desire, consider other alternatives or test before implementing.

Installation

1. Clone the repository

git clone https://github.com/EliasPereirah/SearchAugmentedLLM.git

2. Change directory to the project folder

cd SearchAugmentedLLM

2. Run

composer install

3. Rename .env.example to .env and configure your Google CSE API key and CX ID
4. If you want to use Cohere Rerank, you will need to get an API key from Cohere and add it to the .env file

Features

• Google Search Integration: Leverages the Google Search API to retrieve relevant web pages.
• Content Extraction and Chunking: Extracts text content from web pages and divides it into chunks.
• Relevance Ranking: Re-ranks chunks based on relevance to the initial query using Cohere Rerank model (default: rerank-multilingual-v3.0)
• Contextualized LLM Responses: Delivers the most pertinent information to the LLM, enabling more accurate and informed responses.

API

This project was programmed to be used via REST API, you can use it either on localhost or on an external hosting.

API Parameters

When making an HTTP request to the API, you can pass the following parameters (GET and POST are supported):

Parameter	Description	Required	Default
query	The search query	No
url	URL to extract content	No
time_out	Maximum time (seconds) for a request to each link	No	6
max_results	Maximum number of Google Search results	No	9
max_chunks	Maximum number of chunks to generate	No	100
do_rerank	Rerank results for better quality (requires Cohere API key)	No	true
max_seq	Maximum word length inside a chunk (longer sequences are removed)	No	51
min_char	Minimum number of characters per chunk	No	300
max_char	Maximum number of characters per chunk (must be > min_char + max_seq)	No	450
max_characters_output	Maximum number of characters in the output	No	14000

Update: Now, if you want, instead of passing a search term in the query parameter, you can pass the url you want to get the content from.

Response

You can expect a JSON response with the following keys

text - (string) This is the set of all the chunks in paragraph form which should be passed to LLM.

errors - (array) with error information if there are any.

reranked - (bool) Whether the results were reranked.

snippets - (array) list of snippets returned by Google.

url_references - (object) which references the text by numbering in square brackets

Google CSE API Key

To search using Google, you will need the Google CSE (Custom Search Engine) API Key and CX ID

First, create a custom search here Google CSE Panel

Copy your CX ID -> go to this page on Google Developers and click Get a Key to get your API key. Rename the .env.example file to .env and put your CX and API key in the appropriate variable

Rerank With Cohere

To rerank you will also need to configure a Cohere API key in .env.

Get your Cohere API key here: https://dashboard.cohere.com/api-keys

Security

This project was developed mainly for home use via localhost. If you want to use it on a public hosting, it is recommended to add some restriction layer with login.

License

MIT - This project is licensed under the MIT License.
Please note that this project is currently in beta and is provided "as is" without warranty of any kind.

Acknowledgements

This project leverages the following resources:

Readability PHP library by FiveFilters - https://github.com/fivefilters/readability.php

Cohere API: Used for re-ranking content.

Google CSE for search the web.