SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
What it is
SARA is a research implementation for an ACL 2026 paper that improves Retrieval-Augmented Generation (RAG) by compressing retrieved documents into embeddings. It uses a trained projector to map retriever embeddings into the language model's hidden space at specific token positions, allowing the model to process more documents within a fixed context window.
The problem it solves
Standard RAG approaches are limited by the context window size, forcing a trade-off between the number of retrieved documents and the ability to fit them into the prompt.
What you could use it for
- Comparing standard RAG against compressed embedding RAG on the QASPER benchmark
- Training LoRA adapters and projectors for Mistral-7B-Instruct-v0.2
- Evaluating the impact of context compression on long-document QA tasks
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.