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
Comments
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