CDCK/MoM cites posts as User1, User2 etc -- LLM quality issue?

Falco · August 10, 2026, 4:49pm

Thanks for the report, this PR handles this issue:

github.com/discourse/discourse

DEV: Use a realistic few-shot example in the summarizer agent (#42445)

main ← dev-summarizer-realistic-example

merged 06:41PM - 07 Aug 26 UTC

xfalcox

+14 -3

Previously, the summarizer agent's few-shot example used a two-post toy thread w…hose format didn't match the prompts the summarization strategies actually build, and the system prompt described a post format (`N) username message`) that no strategy emits. This change replaces the example with one mirroring a real captured request (discussion title line, posts inside `<input></input>` tags in the real `(N username said: ...` format) paired with a multi-paragraph cited summary demonstrating `[username]({resource_url}/N)` citations, and fixes the post format description to match what the topic, chat, and gist strategies actually send.

The summary on that topic looks better now.

Topic		Replies	Views
Moderation tool for formatting code with AI Feature ai	10	433	April 3, 2025
Code fences don't seem to deal with C comments Product Support	26	4889	October 6, 2015
HTML comments are also summarized by AI Feature ai , ai-summarize	5	349	July 1, 2026
Pasted content gains erroneous line breaks Bug needs-repro	5	280	August 19, 2024
Catch and educate users posting code without properly formatting it Feature	28	5634	April 11, 2019

CDCK/MoM cites posts as User1, User2 etc -- LLM quality issue?

Related topics