Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
PLBΒ 
posted an update Oct 18
Post
365
πŸ“ˆ Increase the quality of your RAG with a simple Linear Layer! No need to change your embedding model (keep that old OpenAI API).

Introducing EmbeddingAlign RAG, a novel approach to improve Retrieval-Augmented Generation (RAG) systems.

Key highlights:
- Uses a simple linear transformation on existing embeddings
- Boosts hit rate from 89% to 95% on real-world examples
- Minor increase on latency (less than 10ms)
- Works on top of blackbox embedding models (Mistral AI, OpenAI, Cohere,...)
- No dataset needed (just your documents)
- Train easily on CPU

πŸ€— Read the full article here on HF: https://huggingface.co/blog/PLB/embedding-align-rag

Interesting, but how does this approach generalize to arbitrary user query / document domains? Would you need to train a separate network for each domain / dataset?

Β·

As always, there is a trade-off to find between generality and absolute performance. If you have multiple different domaines/dataset with a clear separation, I think it would make sens to train an adapter for each domain.

In this post