CoSearchAgent: A Lightweight Collaborative Search Agent with Large Language Models
Metadata
| Field | Content | |------|------| | Authors | Peiyuan Gong, Jiamian Li, Jiaxin Mao | | Published | 2024-02-09 | | Source | https://arxiv.org/abs/2402.06360 | | Type | Demo paper (arXiv) | | Topic | Conversational / Collaborative Search |
Background and Motivation
Collaborative search supports multiple users working together to accomplish a specific search task. Research has found that designing lightweight collaborative search plugins within instant messaging platforms aligns better with users' collaborative habits. However, due to the complexity of multi-user interaction scenarios, it is challenging to implement a fully functioning lightweight collaborative search system. As a result, previous studies on lightweight collaborative search had to rely on the Wizard of Oz paradigm (simulated system behavior by human operators).
What CoSearchAgent Does
To better support research in collaborative search, the authors propose CoSearchAgent, a lightweight collaborative search agent powered by LLMs:
- Slack plugin: CoSearchAgent is designed as a Slack plugin that supports collaborative search during multi-party conversations on the platform.
- Multi-user understanding: It can understand queries and conversational context in multi-user settings.
- Web search via APIs: It searches the Web for relevant information through APIs.
- Grounded answers: It responds to user queries with answers grounded on the relevant search results.
- Clarification: It can ask clarifying questions when the information needs are unclear.
- Engineering Conversational Search Systems: A Review
- Agentic Conversational Search with Contextualized Reasoning
- ChatRetriever: Adapting LLMs for Robust Retrieval
- ConvGQR: Generative Query Reformulation for Conversational Search
- Original paper: *CoSearchAgent: A Lightweight Collaborative Search Agent with Large Language Models* — arXiv:2402.06360
Significance
The proposed CoSearchAgent is highly flexible and useful for supporting further research on collaborative search, replacing Wizard-of-Oz setups with a real, LLM-driven system. The code and demo video are publicly accessible.