一句话：提出 TenderChat，一个结合查询增强与动态提示的对话式 RAG 框架，用于澳大利亚政府招标信息的检索与解释。
澳大利亚招标流程需要检索大量单份文档以理解需求，现有基于关键词和 ID 的检索方式难以发现新的招标机会；大型语言模型虽能力强，却无法访问实时、结构化的招标信息。检索增强生成（RAG）通过将 LLM 与知识库结合，可实现语义检索与事实性回答并减少幻觉。本文提出 TenderChat：一个新颖的对话式聊天机器人框架，利用查询增强和动态提示高效检索相关信息并生成准确回答。该框架使用多个经提示工程设计的 LLM 对查询进行上下文理解与分类，再据此部署轻量级专用提示用于回答生成，即动态 RAG 方法。系统在领域特定指标上使用 RAGAs、BERT Score 及四个专用 LLM-Judge 进行基准测试，并适配 MultiDoc2Dial 数据集以评估带参考答案的多轮对话。TenderChat 为政府招标的对话式检索与解释提供了实用且可扩展的 RAG 框架。
The Australian tendering process demands searching through many individual documents to understand the requirements. Currently deployed methods of keyword and ID searching are limited in their ability to discover new tender opportunities. While large language models (LLMs) are powerful, they lack access to real-time, structured tender information. Retrieval-Augmented Generation (RAG) offers a solution by combining LLMs with a knowledge base, enabling semantic search and factual responses with reduced hallucinations. In this study, we present TenderChat: a novel conversational chatbot framework that utilises query enhancement and dynamic prompting to retrieve relevant information and generate accurate responses efficiently. TenderChat’s framework utilises multiple prompt-engineered LLMs to understand queries contextually and categorise them for the deployment of lightweight, specialised prompts in response generation. This system is benchmarked using RAGAs, BERT Score, and four specialised LLM-Judges to evaluate domain-specific metrics. Our proposed framework is also adapted for the MultiDoc2Dial dataset, enabling the evaluation of multi-turn conversations with reference answers. TenderChat offers a practical and scalable RAG framework for conversational retrieval and explanation of government tenders, deploying a novel Dynamic RAG approach. Demo: https://youtu.be/EYDL74MYmhY
@inproceedings{paperbot11642,
title = {TenderChat with Dynamic RAG: A Prompt-Adaptive RAG Framework for Australian Government Tender Analysis},
author = {Hayden Michael Fowler and Ruihan Xie and Morteza Saberi and Ali Braytee},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3774748.3787607}
}