ChatRetriever is a large language model (LLM) adapted for conversational dense retrieval, designed to robustly represent complex conversational sessions. The paper introduces a dual-learning approach called Contrastive Session-Masked Instruction Tuning (CSIT) to adapt LLMs for retrieval tasks. This method combines contrastive learning with session-masked instruction tuning, enabling the model to better understand complex conversational contexts and improve retrieval performance. The model is trained on high-quality conversational instruction tuning data and fine-tuned using a contrastive ranking loss function. The results show that ChatRetriever outperforms existing conversational dense retrievers and achieves performance comparable to LLM-based rewriting approaches. Additionally, ChatRetriever demonstrates superior robustness in handling diverse conversational contexts. The paper also presents two robustness evaluation methods to assess the resilience of conversational retrieval approaches. The results highlight ChatRetriever's strong generalization capability in handling diverse conversational search scenarios. The method is evaluated on five conversational search benchmarks, where ChatRetriever achieves significant improvements in retrieval performance. The paper also discusses the limitations of the approach, including efficiency concerns and the need for more effective hard negative mining strategies. Overall, the 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.ChatRetriever is a large language model (LLM) adapted for conversational dense retrieval, designed to robustly represent complex conversational sessions. The paper introduces a dual-learning approach called Contrastive Session-Masked Instruction Tuning (CSIT) to adapt LLMs for retrieval tasks. This method combines contrastive learning with session-masked instruction tuning, enabling the model to better understand complex conversational contexts and improve retrieval performance. The model is trained on high-quality conversational instruction tuning data and fine-tuned using a contrastive ranking loss function. The results show that ChatRetriever outperforms existing conversational dense retrievers and achieves performance comparable to LLM-based rewriting approaches. Additionally, ChatRetriever demonstrates superior robustness in handling diverse conversational contexts. The paper also presents two robustness evaluation methods to assess the resilience of conversational retrieval approaches. The results highlight ChatRetriever's strong generalization capability in handling diverse conversational search scenarios. The method is evaluated on five conversational search benchmarks, where ChatRetriever achieves significant improvements in retrieval performance. The paper also discusses the limitations of the approach, including efficiency concerns and the need for more effective hard negative mining strategies. Overall, the 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.