Most data questions in a company are repeats. "What's our MAU?" gets asked every Monday. Different analysts answer it different ways because each one re-invents the definition.
Verified Knowledge is Saemi's fix: pin the definition once, and every future answer that touches the same metric cites that definition.
How Saemi stores verified knowledge
Saemi tracks three kinds of verified content under the hood. You don't have to think about the distinction day-to-day — the Trust bar under every answer routes you to the right one based on what you're saving.
| Stored as | When you create it | What it's good for |
|---|---|---|
| Schema fact | Trust bar → Add fact → category data-dictionary (the default) |
One-liner truths about your schema — "users.created_at is UTC", "status='cancelled' includes refunds" |
| Knowledge entry | Trust bar → Add fact → pick any other category, OR Save as wiki | Longer definitions, methodology notes, repeatable answers — "Our W2 retention methodology" |
| Verified metric | /knowledge/metrics/new/ (separate page) |
Canonical SQL for a named metric — "MAU = COUNT(DISTINCT user_id) WHERE event='session_start' AND day_window='30d'" |
All three show up in the same Trust bar counter: "3 knowledge entries · 2 schema facts · 1 metric" on every answer they powered.
Save knowledge from a chat
Every assistant answer has a Trust bar at the bottom. Three actions:
- Add fact — opens a textarea + category picker. Default category
data-dictionarysaves a Schema fact; any other category saves a Knowledge entry. Same button, two destinations. - Save as wiki — turn the whole answer into a wiki page (including the SQL and the table). Wiki pages can be promoted to Verified Knowledge later from the page itself.
- View facts — see which existing facts / entries / metrics the answer already used.
Whatever you save shows up the next time someone asks the same question — the answer carries a green Grounded in verified knowledge badge with a link to the source.
How citations work
When Saemi answers, it queries Verified Knowledge first. If a matching fact, metric, or wiki exists, it:
- Uses that definition verbatim (no LLM guessing)
- Cites it inline (you'll see
[knowledge: 42]in the answer markdown, which renders as a clickable badge) - Increments the citation counter on that knowledge entry
You can see citation counts on each knowledge page — "cited 142× this month" surfaces the metrics your team actually cares about.
Organize knowledge
Knowledge lives in categories (Sales, Product, Engineering, Finance, ...). Categories are pre-seeded with a Saemi-recommended starter set; create your own with one click.
Within a category, you can:
- Promote facts to canonical (Saemi will prefer them even when other competing facts exist)
- Mark a fact deprecated (Saemi will warn anyone who tries to cite it)
- Restrict a knowledge entry to specific teams (sales-only metrics don't leak to the product team)
Why this matters
Without Verified Knowledge, every analyst writes their own SQL for MAU. Three analysts produce three slightly-different numbers. The CEO loses trust.
With Verified Knowledge, the first time someone asks, you (the analyst) verify the answer. After that, every Slack mention, every dashboard, every skill that touches MAU runs the same SQL with the same exclusions. One number. Defensible.
The contrast with raw chat
Raw chat (no Verified Knowledge) is fast but flaky — the LLM might guess the wrong column for "active user" in week 3 even if it got it right in week 1.
Verified Knowledge converts that flakiness into a one-time question. After you verify once, the answer is deterministic forever (or until you update the definition).
How to start
The fastest way to bootstrap is to ask Saemi your 5 most-asked weekly questions, then save each answer as wiki + add the 1-2 metric definitions each one used. After a week, ~80% of repeat queries will be cited from knowledge instead of generated.