Effective Dense Retrieval using Only In-Context Examples
RICE gets LLMs to produce stronger retrieval embeddings from a few examples, without retraining the retriever.
The paper introduces RICE, a training-free method that conditions a decoder-only LLM on in-context examples before encoding queries and documents. The authors say those examples give both sides a shared context, improving the dense representations used for retrieval. In their results, RICE substantially improves prompt-based LLM embedding accuracy. The authors also say they are releasing the code. ArXiv · AI/CL/LG's note
The paper introduces RICE, a training-free method that conditions a decoder-only LLM on in-context examples before encoding queries and documents. The authors say those examples give both sides a shared context, improving the dense representations used for retrieval. In their results, RICE substantially improves prompt-based LLM embedding accuracy. The authors also say they are releasing the code. ArXiv · AI/CL/LG's note
score 5