概述
Research evaluates both the answer and the quality of its citation support.
深入探讨
Retrieval-augmented generation (RAG) combines a language model’s parametric knowledge with retrieved non-parametric source material. The original RAG paper evaluates models that retrieve passages from a dense index and condition generation on those passages. ALCE studies end-to-end systems that retrieve evidence and generate answers with citations; it evaluates fluency, correctness and citation quality. Citations improve verifiability only when they support the associated claims. A retrieved document can be irrelevant, incomplete or insufficient, so an application may choose to say the evidence is unavailable rather than guess; this is a design policy, not an automatic property of RAG.
现实世界的实施
Retrieving passages from a document collection before drafting an answer.
Attaching source passages to individual claims in a generated response.
Checking whether cited passages support the claims they accompany.
Evaluating answer quality and citation quality as separate properties.
接地和引文的未来
Test grounded systems with task-matched questions and inspect errors in retrieval, answer correctness and citation support. Report which documents were indexed, which passages were retrieved and how citation quality was measured; do not treat a citation count as a reliability score.
付诸实践
为什么这很重要
- 速度与规模. 语言工作流程可以在不牺牲一致性的情况下更快地移动。
- 交通与覆盖范围. 它扩展了跨语言和沟通方式的访问。
- 更清晰的判决. 团队可以花更多时间进行判断，而自动化则可以处理重复。
留意
- 幻觉的事实可以悄悄地进入报告、支持流程或研究成果。
- 及时的敏感性可能会在类似的请求中产生不一致的结果。
- 如果访问控制薄弱，敏感文本数据可能会暴露。
后续步骤
- 在推出之前定义输出格式、语气和质量标准。
- 当准确性很重要时，请使用可信来源进行地面响应。
- 为高风险输出保留人工审查检查点。
- 跟踪故障模式并定期重新训练提示或工作流程。
不断探索
常见问题
What is Grounding and Citations?
Grounding connects generated answers to retrieved source material, and citations let a reader inspect which evidence is attached to a claim. Research evaluates both the answer and the quality of its citation support.
“扎根”人工智能答案的主要目的是什么？
接地在回答时为模型提供真实的源材料，并要求模型仅根据该材料进行回答，从而使声明可验证。
引文在接地系统中扮演什么角色？
引文充当收据，将每个声明链接到特定的来源段落，以便用户可以验证它。
接地和引用是哪种常见技术的支柱？
RAG 检索相关段落并将其插入提示中，以便模型可以回答并引用真实来源。
If retrieved passages do not support an answer, what conservative design policy can prevent unsupported guessing?
A system can be designed to state that the retrieved sources do not establish an answer rather than guess. This is an application policy, not a guarantee provided by retrieval itself.
为什么不接地气的人工智能答案被认为很难被信任用于高风险用途？
如果没有接地，就没有来源可查，因此无法独立验证说法。