Harden LLM refresh against invented ids; group on teasers #35

Manually merged

marvin8 merged 2 commits from fix/issue-34-refresh-robustness into main

2026-09-12 23:41:36 +00:00

agent-pi commented

2026-09-12 22:50:56 +00:00

Collaborator

Two follow-ups from the post-#23 A/B experiment:

Refresh robustness — the model can hallucinate aggregate ids in a refresh() response; _apply_topics then updated just-deleted aggregates and swallowed the foreign-key errors, producing an empty grouping with a success log. Non-existent ids now degrade to creating a new aggregate, and skipped article links are logged instead of passed silently.
Grouping input reverted to feed teasers — the A/B (recorded on #23) showed full-text prompts fragment clusters with no grouping gain. Grouping runs on teasers again (_CONTENT_PREVIEW_MAX = 300); full text stays in the summarizer teasers and article display.

Closes #34

Two follow-ups from the post-#23 A/B experiment: 1. **Refresh robustness** — the model can hallucinate aggregate ids in a `refresh()` response; `_apply_topics` then updated just-deleted aggregates and swallowed the foreign-key errors, producing an empty grouping with a success log. Non-existent ids now degrade to creating a new aggregate, and skipped article links are logged instead of passed silently. 2. **Grouping input reverted to feed teasers** — the A/B (recorded on #23) showed full-text prompts fragment clusters with no grouping gain. Grouping runs on teasers again (`_CONTENT_PREVIEW_MAX = 300`); full text stays in the summarizer teasers and article display. - [Issue #34 findings + plan comment](https://forge.marvin8.zone/marvin8/yunjin/issues/34) - [A/B results and decision comment on #23](https://forge.marvin8.zone/marvin8/yunjin/issues/23) Closes #34

agent-pi added 2 commits

2026-09-12 22:50:56 +00:00

🐛 Degrade gracefully when the LLM invents aggregate ids 906f4ce596