Alasa Search Engine

Note

This project makes use of embedding and re-ranker Large Language Models in an attempt to semantically understand documents to provide better search results. While the creators are normally extremely Anti-LLM / Anti-Generative AI, they believe that this is an acceptable use of LLMs as nothing is being generated.

This project does make use of open source LLMs that were created by Google and Alibaba Cloud, by default. If you have embedding and/or reranking models trained ethically on public domain data, then we would be happy to change the defaults.

Note

Alasa will never add anything like "search summaries" or anything using LLMs for more than just extracting embeddings from content and calculating the relevance of a result to a user query.

Alasa is an attempt to take techniques used by Generative AI companies to provide a better service to users. Normally, these techniques are used by LLM services to find content relevant to the user's conversation to provide to the LLM for context so that it will hallucinate less. Instead, Alasa takes techniques and uses them to find results for the users directly.

Goals

• Provide superior search results to a classical search engine using semantic search (embeddings / vector search).
• Provide an ElasticSearch-compatible API.
• Use only local LLMs.
• Provide image, audio, and video indexing, along with allowing search by these types of content.

Contributing

Please read the contribution guide.

License

This project is licensed under the Affero General Public License version 3 or later, available in LICENSE. Authors of the project are listed in AUTHORS.md.