Tutorial: Tracing Model Inputs Through Prompts, RAG, and Memory
This tutorial explains how prompts, context, retrieval-augmented generation, and long-term memory contribute different elements to a model’s input. Using synthetic Atlas Notes documents and Open ETF Agent v0.26.0, it shows how to diagnose an incorrect answer by checking whether the task was specified clearly, the required source text reached the request, retrieval selected suitable material, and recalled memory remained applicable. The Open ETF Agent example uses preset actions to simulate model selection, so it does not establish how a real model performs.
The tutorial provides a practical troubleshooting method for locating an incorrect answer’s origin by tracing whether requirements, source text, retrieved content, or remembered information was missing, unsuitable, or outdated. This can make changes to prompts, retrieval, context construction, or memory easier to explain and evaluate.
Prompt、Context、RAG、Memory，模型的输入从哪里来？
你让助手比较两版产品指南。它输出了一张整齐的表，却把旧版限制写到了新版上。你补了一句“认真核对”，结果依旧。此时，究竟该继续改提示词，接知识库，还是给助手加记忆？
这几个选择会改变不同的东西。先把一次模型调用展开，问题就容易定位。
下面继续使用入门篇的 Atlas Notes 合成文档。A 版免费导出支持 CSV，B 版增加 Markdown，两版导出都要求联网。我们的任务是比较导出格式和共同限制。这些是教学材料中的设定。
一次回答之前，模型实际拿到了什么
把本轮上下文想成当前工作台。工作台上可能同时摆着任务说明、产品原文、前一次工具返回，以及需要参考的用户偏好。每样材料有自己的用途。