ETF Agent Maps the RAG Evidence-Delivery Chain
An ETF Agent engineering report organizes RAG into four interacting chains: knowledge construction, online retrieval, evidence consumption and delivery, and evaluation and governance. In the checked private knowledge path, lexical retrieval can use BM25 when the PostgreSQL pg_search extension, a BM25 index, and configuration are available; otherwise it uses PostgreSQL full-text search. Code support alone does not show that any particular query used BM25. The public Open ETF Agent repository offers a smaller teaching implementation with restricted lexical retrieval, original-page readback, and source-linked reporting on synthetic documents, rather than the full private architecture. An RQ-08 evaluation fixed three source versions to one candidate set, compared the ranking of 306 real indexed chunks, and observed actual Knowledge calls. However, the chunks lacked complete manual relevance labels, and some reads confirmed only that a location was found. The results therefore do not establish improved real Recall or end-to-end financial-answer accuracy.
The report provides a concrete account of how RAG can preserve source identity and versions, apply authorization before ranking, distinguish retrieval from actual model consumption, and bind citations to delivered evidence. It also separates those engineering controls from answer-quality gains that were not demonstrated.
` 同步校注（2026-11-04）：文中涉及撤权、来源消费和恢复的历史案例，应与当前 Run 授权生命周期区分。当前实现中，Host 已授予的权限在同一 Run 及其恢复中持续有效，不因外部权限表变化重新撤销；新 Run 独立准入。账户、scope、来源范围、内容身份、预算和显式取消仍按原合同执行。历史撤权复现用于解释当时的问题，不代表当前 Run 内重新授权机制。
同一个问题，一套检索系统可能给出三份看似相关的基金资料：2021 年的旧招募说明书、2023 年的更新公告、2025 年的后续招募说明书。用户问的是：“2023 年到底改了什么？后来的材料能否解释当时的变更？”
把三份资料送进模型，答案不一定更好。最新的文件未必是历史事件的直接依据；命中“现金替代”几个字，不代表找到了适用于这次变更的条款；即使引用链接正确，模型也可能把“后来出现的字段”写成“当时修改的原因”。
#ETF Agent#Open ETF Agent#PostgreSQL#PostgreSQL pg_search#BM25#PostgreSQL FTS#RRF#TEI