Break The Vector Search Dependency for Truly Robust RAG Systems

Author(s): Thuwarakesh Murallie

Originally published on Towards AI.

Why your RAGs need more than just semantic search to succeed.Photo by Marek Piwnicki on Unsplash

Vector search is undoubtedly the most remarkable retrieval technique to date. But this can be an overstatement.

Before vector stores, we searched for information in various ways. Some of our favorites included full-text search, trigram similarity, and BM25. Most popular databases, like Postgres, have built-in capabilities to perform these searches.

But they had a common weakness — their semantic understanding is poor.

Vector search isn’t new, but it was once considered computationally costly. By today’s standards, the cost isn’t too high. However, vector search is worth the cost for its contextual awareness and semantic similarity.

It had almost become the de facto for RAGs. A lot of vector databases have become very popular.

But we started to rethink its usefulness once again. Is vector search worth it?

The answer is both yes and no!

Most RAGs are built on this stack; why would you redo it every time?

ai.gopubby.com

Vector databases are fast and impressive in retrieving information that is not obvious, yet relevant.

But there are gray areas where vector search doesn’t do a good job.

Firstly, vector search performs poorly when uncommon abbreviations are used. The embedding model is trained with a large corpus of public data, so it can handle standard abbreviations, such… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI