Title: RCEM: A Robust Conversational Search Embedder with Integrated Query Rewriting

Abstract:

As retrieval-augmented generation (RAG) systems grow in prominence, conversational search has emerged as a critical component, allowing users to engage with AI assistants via multi-turn dialogues featuring context-sensitive queries. To address this, we introduce RCEM, a conversational dense retrieval architecture that transfers the query reformulation prowess of Large Language Models (LLMs) into an embedding model. This approach facilitates context-aware retrieval during inference without the need for explicit query rewriting steps.

In contrast to earlier conversational dense retrieval methods that focus on learning direct mappings from conversation turns to documents, RCEM aligns embeddings of conversational queries with those of their rewritten counterparts. This alignment strategy significantly enhances robustness when facing distributional shifts. Furthermore, RCEM eliminates the necessity for training data consisting of conversational query-to-document relevance judgments, which are typically costly to acquire and challenging to maintain at a high quality.

Comprehensive evaluations on the QReCC, TopiOCQA, and TREC CAsT benchmarks reveal that RCEM consistently surpasses robust conversational retrieval baselines. Notably, it delivers substantial performance improvements under distributional shift conditions, including a Recall@10 increase of up to 20%. By endowing the base embedding model with conversational query rewriting abilities while maintaining its original retrieval functions, RCEM allows a single model to encode both standalone and conversational queries. This capability enables search against existing document indexes without the requirement to rebuild the retrieval database.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC