AI Platform overview
How Orion, the Context Studio AI Hub, Domain Skills, and Orion Memory fit together into one governed AI layer.
When to use it: When you are new to Celiq's AI and want to understand how the pieces connect before diving into any single feature.
Where to find it: Orion lives in the left sidebar and inside Discover and Mosaics. Its knowledge and quality controls live in Context Studio (left sidebar → Context Studio, route
/context-studio).Who can use it: Anyone can ask Orion (Tracer and above). The Context Studio controls that shape Orion are for Weaver and Luminary.
Celiq's AI is not one feature — it is a layer woven through the product, backed by your governed semantic layer. This page ties the four pieces together so you know which one to reach for: Orion (the assistant you talk to), the Context Studio AI Hub (where you watch and measure that assistant), Domain Skills (what you teach it), and Orion Memory (what it remembers). Each piece has its own canonical page — this overview is the map, not a duplicate.
Overview
Think of the AI platform as one assistant with a brain you can inspect and improve:
| Piece | What it is | Where it lives | Who shapes it |
|---|---|---|---|
| Orion | The AI analyst — ask questions in plain English, generate CML, explain SQL | Sidebar, Discover, Mosaics | Tracer+ uses it |
| AI Hub | Real-time view of AI activity across the product, with accuracy and a corrections queue | Context Studio → AI Hub | Weaver, Luminary |
| Domain Skills | Plain-English business context Orion loads before answering in a domain | Context Studio → Domain Skills | Weaver, Luminary |
| Orion Memory | Facts, preferences, and instructions Orion has learned about your workspace | Context Studio → Orion Memory | Weaver, Luminary |
When you ask Orion a question, it pulls your semantic layer (nodes, metrics, governed reveals), loads the relevant Domain Skill, applies anything in Orion Memory, answers, and then records that exchange so the AI Hub can show it to you. Improving an answer means editing the skill or the memory — and the AI Hub is where you see whether your change worked.
When to use it
- You are onboarding and want to understand Celiq's AI before reading the deep-dive pages.
- An Orion answer was wrong and you are not sure whether to fix it in Domain Skills, Orion Memory, or by certifying a reveal.
- You need to explain to a teammate how Celiq's AI stays governed instead of hallucinating.
- You are a Weaver or Luminary deciding where to invest to make Orion smarter about your business.
Concepts
Orion — the AI analyst built into Celiq. It answers natural-language questions using your semantic layer, can translate questions into a governed Reveal, explain generated SQL, and draft CML. See Orion. Governed reveal — a named, modelled query in a domain. When Orion answers from a reveal, the answer uses your canonical metric definitions rather than ad-hoc SQL. Certified reveals are trusted most. Confidence tier — every Orion answer carries a confidence that reflects its source, not its wording: a certified reveal answers at high confidence, an uncertified reveal at medium, and a raw SQL fallback at low. Domain Skill — plain-English context for one business area (its rules, date logic, and data quirks) that Orion loads before answering in that domain. See Domain Skills. Orion Memory — workspace-level facts, preferences, instructions, and context Orion has learned and reuses across sessions. See Orion Memory. AI Hub — the live activity and quality surface in Context Studio. It shows questions answered, accuracy and eval pass rate, a source breakdown across Discover / Mosaics / Forge / Orion, and a pending corrections queue. See AI Hub. Skill Registry — the catalogue of everything Orion can invoke: compute skills (live tools) and knowledge skills (authored context). Every AI surface shares the same registry. See Skill Registry.Getting started
Two surfaces matter:
- Orion is always reachable — open it from the left sidebar, or from the Ask Orion panel inside Discover and any Mosaic. Asking questions requires the Tracer role or above.
- Context Studio is where you shape Orion. Open it from the left sidebar (route
/context-studio). The route admits Weaver and Luminary; some panels, such as AI Operations and certifying reveals, are Luminary-only.
Before Orion can answer, a workspace AI key must be configured. If it is not, Orion returns a setup prompt rather than an error and points you to Admin Console → AI Settings. A Luminary configures the key once for the whole workspace.
Step-by-step
Ask Orion a question
Open Orion from the sidebar or the Ask Orion panel in Discover. Ask in plain English, e.g. "Show me revenue by country last 90 days." Orion picks the right node, dimensions, metrics, and filters and runs it.
Read the confidence signal
Check the confidence on the answer. High means it used a certified reveal; medium means an uncertified reveal; low means a raw-SQL fallback. Low or medium is your cue that the semantic layer could be tightened.
Open Context Studio to see the bigger picture
Go to Context Studio → AI Hub. The overview shows questions answered, accuracy, eval pass rate, and any corrections pending. The event feed shows AI activity across Discover, Mosaics, Forge, and Orion.
Teach Orion where it was weak
If a domain's answers are off, edit its plain-English context in Domain Skills. If Orion should always know a specific fact or preference, add it in Orion Memory. Both take effect on the next question.
Measure the improvement
Run Accuracy Evals to score Orion against a question set, then watch the pass rate in the AI Hub. The loop is: spot a weak answer → fix the skill or memory → re-run evals → confirm the pass rate moved.
Examples
A worked loop, from a wrong answer to a durable fix.
You ask Orion "What was revenue last month?" and the number looks too high — it included cancelled orders. That is a Domain Skill problem: the orders domain has not told Orion how to treat cancellations. Open Context Studio → Domain Skills, select the orders domain, and add a rule in plain English:
Revenue excludes cancelled and refunded orders.
Order counts exclude rows where status = 'Cancelled'.
Our fiscal month starts on the 1st in the workspace timezone (UTC).Save. Orion picks up the change on its next question. If this is a fact Orion should apply everywhere — not just one domain — add it to Orion Memory instead, as a fact or instruction:
type: instruction
label: cancelled_orders
value: Always exclude status = 'Cancelled' from order counts and revenue.Then add the original question to Accuracy Evals with the correct expected answer, run the evals, and confirm the pass rate in the AI Hub went up. The fix is now both applied and protected against regression.
Best practices
- Match the fix to the scope. A rule that is true for one business area belongs in that Domain Skill; a rule true everywhere belongs in Orion Memory.
- Certify your key reveals. Certified reveals are what Orion trusts most and what lets answers reach high confidence — they are the single highest-value investment in answer quality.
- Use the AI Hub as your scoreboard. Watch accuracy, eval pass rate, and the corrections queue rather than judging the AI from one answer.
- Close the loop with evals. After every meaningful skill or memory change, re-run Accuracy Evals so improvements are measured, not assumed.
- Be specific in skills. "Revenue excludes refunds and chargebacks" teaches Orion far more than "Revenue is total sales."
Tips
The confidence tier is your map. A run of medium answers in one area means an uncertified reveal is doing the work — certify it. A run of low answers means there is no governed reveal at all, so Orion is falling back to raw SQL — build the reveal in Forge.
Common mistakes
Editing data to fix an answer. If Orion returns the wrong number, fix the metric definition, the domain skill, or the memory — never paper over it by changing rows.
Expecting Orion to read your rows. Orion reasons about your semantic layer and metadata, not the values inside individual rows. Give it the rules; it applies them.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| Orion returns a setup prompt instead of an answer | No workspace AI key is configured | A Luminary sets the key in Admin Console → AI Settings |
| "Orion is not configured" / request blocked | The AI provider key is missing or invalid | Re-check the key in Admin Console → AI Settings |
| Answers are consistently medium confidence | The reveal in use is uncertified | Certify the reveal so answers can reach high confidence |
| Answers are low confidence or use raw SQL | No governed reveal exists for that metric | Build the reveal in Forge, then re-ask |
| "Too many Orion requests" | The hourly per-user rate limit was hit | Wait and retry; Luminary accounts have a higher limit |
| Orion query limit reached this month | The plan's Orion query quota is used up | Upgrade the plan, or wait for the monthly reset |
| A fix to a skill did not change the answer | The skill is saved but the question routed to a different domain | Confirm the domain, and check Orion Memory for an overriding instruction |
Related pages
- Orion — the AI analyst in depth
- AI Hub — real-time AI activity, accuracy, and corrections
- Domain Skills — teach Orion each business area
- Orion Memory — what Orion remembers across sessions
- Skill Registry — every compute and knowledge skill Orion can invoke
- Accuracy Evaluations — score and track Orion's accuracy