ChatRetriever: Adapting LLMs for Generalized and Robust Conversational Dense Retrieval
- Paper: ChatRetriever: Adapting Large Language Models for Generalized and Robust Conversational Dense Retrieval
- Authors: Kelong Mao, Chenlong Deng, Haonan Chen, Fengran Mo, Zheng Liu, Tetsuya Sakai, et al. (7 authors)
- Published: 2024-04-21
- Link: https://arxiv.org/abs/2404.13556
- Topic area: Conversational Search
- Evaluated on five conversational search benchmarks.
- Substantially outperforms existing conversational dense retrievers.
- Achieves state-of-the-art performance on par with LLM-based rewriting approaches.
- Shows superior robustness across diverse conversational contexts.
- Neural IR has evolved from BM25 to BERT cross-encoders, bi-encoder dense retrieval, late interaction, generative retrieval, and now LLM-based agents. ChatRetriever sits in the dense-retrieval line, using LLMs as session encoders.
- Pipeline approaches (rewrite → retrieve) work well but add latency and can propagate rewriting errors; adapting the retriever directly to sessions is an appealing alternative.
- Engineering considerations for deploying such systems: latency budgets and retrieval-step control, index update frequency, embedding version management, cache strategies, security concerns (corpus poisoning, bias), and per-query cost.
- Open problems noted in this research area include evaluation reliability, latency and cost constraints, hallucination and safety, and multilingual/multimodal extension.
- A Survey of Conversational Search, Sep 2025, ACM (survey, ACM)
- Engineering Conversational Search Systems: A Review of Applications
- CoSearchAgent: A Lightweight Collaborative Search Agent with LLMs (arXiv:2402.06360)
- ConvGQR: Generative Query Reformulation for Conversational Search (arXiv:2305.15645)
- Agentic Conversational Search with Contextualized Reasoning via Reinforcement Learning
- CTR-Guided Generative Query Suggestion in Conversational Search (EMNLP)
- Original paper: ChatRetriever: Adapting Large Language Models for Generalized and Robust Conversational Dense Retrieval. https://arxiv.org/abs/2404.13556
Summary
Conversational search requires accurately interpreting user intent from complex multi-turn contexts. ChatRetriever inherits the strong generalization capability of large language models to robustly represent complex conversational sessions for dense retrieval. Instead of relying on a separate query-rewriting stage, the LLM itself is adapted to encode multi-turn sessions directly.
Method
The paper proposes a dual-learning approach:
1. Contrastive learning — adapts the LLM as a dense retriever. 2. Masked instruction tuning — enhances complex session understanding using high-quality conversational instruction tuning data.
This design lets a single model both retrieve and deeply understand conversational context, avoiding the latency and error-propagation costs of pipeline-based rewriting.
Results
Original Abstract
> Conversational search requires accurate interpretation of user intent from complex multi-turn contexts. This paper presents ChatRetriever, which inherits the strong generalization capability of large language models to robustly represent complex conversational sessions for dense retrieval. To achieve this, we propose a simple and effective dual-learning approach that adapts LLM for retrieval via contrastive learning while enhancing the complex session understanding through masked instruction tuning on high-quality conversational instruction tuning data. Extensive experiments on five conversational search benchmarks demonstrate that ChatRetriever substantially outperforms existing conversational dense retrievers, achieving state-of-the-art performance on par with LLM-based rewriting approaches. Furthermore, ChatRetriever exhibits superior robustness in handling diverse conversational contexts. Our work highlights the potential of adapting LLMs for retrieval with complex inputs like conversational search sessions and proposes an effective approach to advance this research direction.