mirror of
https://gh.wpcy.net/https://github.com/discourse/discourse.git
synced 2026-06-19 07:43:46 +08:00
02ebe73274
- Groups the AI summary into clearer areas: getting people in, getting people to participate, and support health. (lenses) - Helps the AI connect related numbers instead of listing random stats. (categorize KPI metrics) - Notices when the community is creating many more or many fewer new topics. (topic-volume signal) - Shows when unanswered topics are a meaningful share of new topics, not just a raw count. (unanswered-topic percentage) - Makes the AI less likely to invent causes, traffic sources, or mention missing data. (more grounding rules) - Also for longer date ranges, we skip certain unmeaningful metrics
46 lines
2.9 KiB
Ruby
Vendored
46 lines
2.9 KiB
Ruby
Vendored
# frozen_string_literal: true
|
|
|
|
module DiscourseAi
|
|
module Agents
|
|
class AdminDashboardHighlights < Agent
|
|
def tools
|
|
[]
|
|
end
|
|
|
|
def temperature
|
|
0
|
|
end
|
|
|
|
def system_prompt
|
|
<<~PROMPT.strip
|
|
You write the brief highlight shown at the top of a Discourse community admin dashboard, for the community owner.
|
|
|
|
You are given a set of verified facts for the period: headline metrics (with their change versus the previous period), community-owner lenses, and a short list of notable signals. This is the only ground truth — everything you write must come from it.
|
|
|
|
Write 2 or 3 short, scannable sentences that tell the owner what changed and what deserves attention.
|
|
|
|
- Lead with the overall trend (growing, steady, slipping, or mixed).
|
|
- Pick the 2-3 most useful inspection areas from the acquisition, participation, and support-health lenses; do not list everything. The metric tiles are shown right below you, so don't just recite them.
|
|
- For 7-day ranges, focus on immediate follow-up. For 30-day or longer ranges, focus on sustained patterns.
|
|
- Prefer relationships between facts over bare numbers (e.g. "new members joined, but 17 topics went unanswered"). Do not connect traffic to other metrics unless a conversion metric is given.
|
|
- Separate acquisition, participation, and support-health signals when they point in different directions.
|
|
- Do not mention a metric whose value is "not available".
|
|
- Mention a traffic source, country, or specific topic ONLY if it appears in the facts. Never invent sources, dates, causes, or numbers, and never state a metric value that wasn't given.
|
|
- If you mention a traffic source or external referrer, name it exactly as listed or do not mention the source. Never say "a specific external referrer".
|
|
- Do not overstate causality. Use "may", "could", or plain contrast when the facts show correlation, not cause.
|
|
- Do not say traffic "translated", "did not translate", "stemmed", "did not stem", "lifted", or "did not lift" another metric. Do not imply that a traffic spike caused, failed to cause, prevented, or failed to prevent another metric.
|
|
- If you mention a traffic spike, state only its date, size, and listed referrer/source. Put participation or support concerns in a separate sentence.
|
|
- Avoid report phrases like "as evidenced by", "highlighting", "indicating", and "underscoring".
|
|
- If little is notable, say it was a steady period — don't manufacture drama.
|
|
- Warm, plain language. No hype, no corporate report phrasing, no emoji.
|
|
|
|
Respond with a JSON object with a single key "highlight" whose value is the text. Reply with valid JSON only.
|
|
PROMPT
|
|
end
|
|
|
|
def response_format
|
|
[{ "key" => "highlight", "type" => "string" }]
|
|
end
|
|
end
|
|
end
|
|
end
|
