Semantic search system

About

This is a simple example of a semantic search system that implements both Elasticsearch and FAISS.

The reasoning behind this is that ES's support for text/sentence embeddings is fairly new, limited, and a field of ongoing work (Ref: Text similary search with vector fields). There are plugins that can implement approximate nearest neighbor (ANN) solutions in ES as described here: Scalable Semantic Vector Search with Elasticsearch. However, I wanted to keep ES's native search possibilities but have an ANN solution separately.

How it works

front is an API that communicates with ES and semantic server which is an API that represents FAISS.

In this example there are three kinds of queries:

• native - Text ES queries
• semantic - Semantic query that returns first top N hits
• semantic range - Semantic query that returns all articles in a defined range of similarity score

semantic server implements XLM-RoBERTa (optimized for sentence embeddings) which calculates query embeddings and searches FAISS index. In order to relate embeddings from FAISS with articles from ES, semantic server contains article uuids that are stored during the creation of FAISS index. In order for XLM-RoBERTa to work on large text queries, text is divided into chunks which are processed separately and combined into one embedding vector.