Effects of Retrieval Instructions in Embedding Models
What it is
This paper investigates how retrieval instructions affect the representations of queries in asymmetric retrieval tasks. It demonstrates that current embedding models often fail to follow simple instructions when query-side distractors are present, attributing this to their training setup. The authors show that fine-tuning with added query-side distractors leads to substantial improvements in instruction following with minimal impact on other tasks.
The problem it solves
Current embedding models struggle to reliably follow detailed retrieval instructions, particularly when query-side distractors are included, leading to poor retrieval performance.
What you could use it for
- Improving the robustness of embedding models for retrieval-augmented generation (RAG) systems
- Designing better evaluation benchmarks for instruction-following in embedding models
- Developing fine-tuning strategies to mitigate sensitivity to query-side distractors
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.