This end-to-end snapshot was produced by the current prompt-native Co-STORM
preview on 2026-07-11. It shows how one warm start can continue through a
user-steered sequence and then close with a final file artifact. The visible
participants are simulated research roles, not independently running agents or
the upstream CoStormRunner.
The checked-in Pages snapshot uses a descriptive publication filename. A
standard Co-STORM file-producing run writes this requested report as
co_storm_report.html unless the user explicitly chooses another name.
The report covers production RAG architecture, Chunking, Hybrid Retrieval, Reranker design, advanced RAG boundaries, evaluation, security, and an implementation roadmap. It cites 20 papers or official technical sources.
Use the storm skill in Co-STORM mode to research RAG technology.
The run did not use a fixed number of rounds. Each numeric reply selected one of the visible choices offered by the Moderator.
| Stage | User input | Selected branch |
|---|---|---|
| Warm start | Initial prompt | RAG foundations, retrieval architectures, production evaluation, and the initial mind map |
| Round 2 | 1 |
Production-grade RAG architecture |
| Round 3 | 2 |
Chunking, Hybrid Retrieval, and Reranker |
| Round 4 | 1 |
Golden test sets and launch metrics |
| Round 5 | 1 |
Executable evaluation contracts |
| File-producing turn | 1 |
Persist the evaluation kit and validate it |
| Conclusion | 生成报告 |
Synthesize the selected branches into the final HTML report |
The concluding user message was deliberately short:
生成报告
The preceding turn had already explicitly authorized persistent file output, so the instruction closed the roundtable and produced the HTML report. In a fresh conversation, make the file intent explicit, for example:
Conclude the Co-STORM roundtable and save the final report as HTML.
Without file authorization, the conversation-first mode should return the report in chat instead of writing it to disk.