Orion

Orion is the AI layer woven through Celiq. It is not a separate product — it is present wherever you are working. Ask a question in Discover, generate a node in Forge, or get a daily briefing in your inbox. Orion handles it.

What is Orion

Orion is a large-language-model assistant that understands your semantic layer. Because it has access to your CML — your field definitions, metric formulas, and join graph — it gives answers that are grounded in your actual data model, not generic SQL guesses.

Orion can:

• Translate natural language into a Reveal (select fields, apply filters, pick a visualisation)
• Explain the SQL behind any query in plain English
• Generate a node file in Forge from a description or table name
• Describe a metric — explain what a measure calculates and what caveats apply
• Identify anomalies in your data and surface them proactively
• Send a morning briefing with changes in your key metrics since yesterday

Orion in Discover

Ask a question

Click the Orion button (sparkle icon) at the top of Discover, or press Cmd/Ctrl + K. Type a question in plain English:

• "Show me revenue by country last 90 days"
• "What were the top 10 customers by order value in Q1?"
• "Which product categories had the biggest month-on-month drop?"

Orion translates your question into a Reveal — picking the right node, dimensions, measures, filters, and date range — and runs it immediately. You can then edit the Reveal manually or ask a follow-up question.

Explain SQL

Every Reveal in Discover has an Explain button (in the ⋯ menu). Click it to see:

1. The raw SQL Celiq generated
2. A plain-English explanation of what the SQL does
3. Notes on any joins or aggregations that might affect the result

This is especially useful when onboarding new analysts who want to understand what Celiq is actually querying.

Ask follow-up questions

After running a Reveal, Orion suggests three follow-up questions based on the results. Click any suggestion to run it instantly. You can also type your own follow-up in the Orion bar — Orion maintains context from the previous query.

Proactive anomaly detection

After Orion answers your question, it silently checks 2–3 related metrics for unusual movement. If anything is outside the normal range (±20% from the 4-week rolling average), Orion flags it.

You ask: "What's our revenue this week?"

Orion answers: "Revenue was £124k — up 18% WoW."

Then adds: "Worth flagging: Conversion rate ↓ 25% vs 4-week average (1.8% vs baseline 2.4%)."

This happens automatically. You do not need to enable it. To adjust sensitivity or disable it, go to Lumen → Orion settings → Anomaly detection.

Suggested actions

After answering questions about metric changes, Orion suggests what to investigate or communicate next.

For a revenue drop, Orion might suggest:

• Check channel breakdown → click to run the query instantly
• Check product mix → click to see which products drove the change
• Alert team → click to draft a message with the key numbers filled in

Suggested actions appear below the answer. Click any suggestion to run it as a follow-up query, or use the draft message template to copy a pre-filled Slack or email message.

When Orion can't find a governed reveal

If you ask about a metric that has no governed reveal, Orion stops and explains the situation clearly rather than guessing with raw SQL:

"I don't have a governed reveal for campaign revenue yet.

Here's what I can do:

Option A — Approximate now: Run an ad-hoc SQL query. Fast, but won't use your governed revenue definition.

Option B — Build it properly: I can guide you to create a reveal in Forge (~2 minutes)."

Orion waits for you to choose. This protects you from getting numbers that look right but use the wrong metric definition.

Orion in Forge

Generate a node

Click the Orion button in the Forge toolbar. Describe the table you want to model:

"Generate a node for the analytics.orders table. Include order_id, customer_id, revenue, status, and created_at. Add row-level security by customer_region."

Orion reads your schema (via the active connection) and writes a complete CML node file with:

• Dimensions with sensible labels and descriptions
• At least one measure (usually a count and a sum)
• A Data Key block if you asked for RLS
• Comments on any fields it made assumptions about

Review the output, make any edits, then validate and commit.

Describe a metric

Right-click any measure in the Forge editor and select Ask Orion: Describe this metric. Orion explains:

• What the SQL calculates
• Which tables and joins are involved
• Any edge cases (e.g. "this count is distinct — duplicate order_ids are collapsed")
• Suggested documentation text you can paste into the description: field

Morning briefing

The morning briefing is an Orion-generated email or Slack message delivered once a day (default: 8am your local time). It contains:

• The top 3–5 metrics from your pinned Mosaics, with a delta since yesterday
• Any anomalies Orion detected overnight (metric moves more than 2 standard deviations from its rolling average)
• Suggested follow-up questions for each anomaly

To configure it:

1. Open your Profile → Notifications.
2. Enable Morning briefing.
3. Choose delivery channel: email, Slack DM, or Teams DM.
4. Select which Mosaics to include.

Role-based briefings

Celiq sends a personalised morning briefing to each user based on their role — CEO, ops, marketing, product, or finance. Each role receives different metrics:

Briefings are generated by Orion at 7am in your workspace timezone.

To configure:

1. Go to Lumen → Orion settings → Briefings.
2. Set your role and preferred delivery channel (Slack or email).
3. Toggle briefings on.

Briefings include anomaly flags for your role's metrics. If a metric moves more than the escalation threshold (configurable per role), an immediate alert is sent instead of waiting for the next morning briefing.

How Orion knows what you want

Orion automatically detects whether you're asking a question or asking it to build something.

• "What's our revenue this week?"
• "Why did conversion drop on mobile?"
• "Which channel drives the most orders?"

In query mode, Orion uses your governed reveals, raw SQL, or workspace memory — whatever gets you the most accurate answer fastest.

• "Build me a revenue dashboard"
• "Create a mosaic for our ecommerce data"
• "Add a tile showing orders by channel"

In build mode, Orion only uses governed reveals — never raw SQL. This ensures every tile uses your canonical metric definitions. If a reveal doesn't exist for a metric you want on a dashboard, Orion tells you and offers to create one in Forge.

The mode switch happens automatically based on the intent in your message. You never need to specify it.

How Orion knows your business rules

Orion loads domain-specific knowledge before answering questions. For an orders question, it loads rules like:

• Revenue = SUM(sale_price), not SUM(cost)
• Exclude cancelled orders from order counts
• AOV = revenue / COUNT(DISTINCT order_id), not AVG(sale_price)

These rules are stored in domain skills files — one per business domain (orders, marketing, products, customers, finance, ops). Celiq ships with a standard set, and they can be customised per workspace in Lumen → Orion → Skills.

When you ask a question, Orion identifies the relevant domain and loads its skill file before generating any SQL or selecting any reveal. This is why Orion gets metric calculations right on the first try for well-defined domains, rather than guessing at what "revenue" means.

What Orion remembers

Orion maintains a workspace memory that persists across sessions. It stores:

• Your canonical metric definitions and the gotchas that come with them
• Data quality issues discovered during past queries (e.g. "always exclude status = 'Cancelled' from order counts")
• Corrections you've made ("you told me last session that AOV excludes refunds")
• Your preferred analysis grain (weekly vs daily)

Memory is loaded silently at the start of every session. It does not add to your query time — it is retrieved once when Orion starts, not on each message.

View and edit what Orion remembers at Lumen → Orion → Memory. You can delete individual memory entries or clear all memory for a fresh start.

Provenance footer

Every Orion response shows three trust signals at the bottom:

• Source — semantic layer (governed reveal) or raw SQL
• Freshness — the date of the most recent data in the result
• Confidence — high (governed reveal) or medium (raw SQL)

High confidence means the answer used a metric definition you or your team wrote in CML. Medium confidence means Orion approximated the metric from your schema — verify the definition is correct for your business.

Anomaly detection

After returning an answer, Orion asynchronously checks 2–3 related metrics in the background. If any metric is more than ±20% from its 4-week rolling average, Orion flags it as an anomaly with a suggested follow-up question.

Anomaly checks never delay the primary response — they run after the answer is shown and surface as a follow-up notification if something is found.

Follow-up suggestions

After every analytical response, Orion suggests 2–3 follow-up questions as clickable buttons. Click any suggestion to run it instantly. Orion maintains context from the previous query, so follow-ups build on what you just asked.

Action suggestions

After explaining an insight, Orion suggests what to do next:

• Investigate — a button that sends a specific drill-down query directly
• Draft message — copy-ready text summarising the insight to send to your team

Example: "Revenue dropped 23% last week. [Investigate by channel →] [Copy summary for Slack →]"

Graceful degradation

If no governed reveal exists for your question, Orion stops and offers two paths rather than guessing silently:

• Option A — run an approximate query using raw SQL (with a medium-confidence caveat)
• Option B — create a reveal in Forge first, then re-run the question with high confidence

Orion will never substitute a different metric than the one you asked for.

Role-based briefings

The morning briefing is personalised per role:

• CEO / founder — revenue, growth, and key conversion metrics
• Operations — fulfillment rates, returns, and SLA metrics
• Marketing — channel performance, CAC, and campaign metrics
• Finance — MRR, ARR, burn rate, and cohort metrics

Set your briefing role at Lumen → Profile → Briefing role.

Limitations

Orion is powerful but not perfect. Be aware of these limitations:

• Orion does not write your SQL for you. It generates CML and Reveals, which Celiq then compiles to SQL. It does not execute arbitrary SQL directly.
• Orion can hallucinate field names if your schema has ambiguous or abbreviated column names. Always validate generated CML before deploying.
• Orion has no access to your actual row data. It reasons about metadata (column names, types, cardinalities) and your CML definitions — it cannot read the values in your rows.
• Follow-up context is session-scoped. Closing the Discover tab clears Orion's context. It does not remember previous sessions.
• Languages: Orion accepts questions in any language but returns answers and CML in English. Multi-language label support is on the roadmap.

Proactive scanning

Orion's proactive scanning is automatic. Every 6 hours, Orion compares the current week to the prior week across all your active reveals. Findings with more than 15% change are surfaced.

What gets surfaced

• Single metric changes — revenue up 34%, orders down 12%
• Opposing metric patterns — revenue up but new customers down
• Correlation signals — traffic up but conversion down suggests a quality issue

Investigating a finding

Click Investigate with Orion → on any finding card to open the Orion panel with the question pre-filled. Orion runs root cause analysis on the finding immediately.

Sensitivity

Adjust when findings are surfaced: Lumen → Orion → Scanning sensitivity. Options: Low (30% change), Medium (15%), High (5%).

{/ Screenshot: Homepage with "Orion noticed" finding cards and percentages /}

Auto chart selection

Celiq picks the chart type based on:

Overriding

Use the chart switcher below any Orion response to change the chart type manually. Your choice is remembered for that session. When auto-selected, a small auto label appears next to the chart type.

{/ Screenshot: Orion response with auto-selected bar chart and chart type switcher visible /}

Decomposition tree

When you ask Orion to explain why a metric changed, it builds a decomposition tree: the top-level metric splits into the most impactful dimension, then each segment splits again.

Each node shows the value and its percentage of the parent. Click any node to drill deeper into that segment.

Example questions that trigger the tree

• "Why did revenue drop this week?"
• "Break down orders by what's driving them"
• "What explains the conversion rate change?"

{/ Screenshot: Decomposition tree for revenue breakdown showing nested dimension splits /}

AI Query Autocomplete

When you open the Orion bar, a dropdown of suggestions appears immediately — before you have typed anything. As you type, the list narrows to match.

How starter suggestions work

When the input is empty, Orion shows starter suggestions — a short list of high-value questions tailored to your workspace. They are generated from:

• Your governed reveals — so every starter routes to the semantic layer and returns a high-confidence answer.
• Your most-asked questions — recent popular queries across the workspace.
• Your role — a marketer sees channel and CAC starters; finance sees MRR and margin starters.

Example starters for an ecommerce workspace:

• "Show me revenue by country last 90 days"
• "What's our conversion rate this week vs last week?"
• "Top 10 products by revenue this month"

Because starters are built from reveals that exist, they almost always answer from the semantic layer rather than falling back to raw SQL.

As-you-type suggestions

Once you start typing, suggestions update live. Each suggestion shows:

• The full question text
• A small badge indicating the route it will take — semantic layer (governed) or raw SQL (ad-hoc)
• The reveal it will use, when applicable

Selecting a suggestion fills the bar with the full question. Press Enter to run it, or keep editing first.

Keyboard shortcuts in the suggestion dropdown

The dropdown is fully keyboard-navigable, so you never need to reach for the mouse:

If no suggestion is highlighted, Enter runs whatever you have typed.

Limitations

• Suggestions are generated from your current reveals and recent query history. A brand-new workspace with no reveals shows generic starters until you build your semantic layer.
• Suggestions reflect your access level — you will never be suggested a question over a field hidden from you by a Data Gate.

Workspace Intelligence Score

Every question you ask, every correction you make, and every reveal you certify teaches Orion something specific to your business. The Workspace Intelligence Score turns that accumulated knowledge into a single number, so you can track how Orion grows from a generic assistant into one that knows your business.

The homepage widget

The Intelligence widget on the homepage shows your current score, level, and a one-line summary:

🧠 Orion Intelligence                    [Trusted]

  73 / 100
  ████████████████████░░░░░░

  Orion has learned 47 things about your business

  [See what Orion knows →]

Click See what Orion knows to open the breakdown.

How the score is calculated

The score is the sum of five components, each capped so no single source dominates:

The components add up to a maximum of 100. The score is recalculated periodically (it does not need to be real-time) and cached, so opening the widget is instant.

Levels

Your score maps to one of four levels:

The level is a quick signal of how much you can lean on Orion's governed answers without re-checking them.

What "Orion knows" means

The breakdown panel lists the component scores with mini bars, plus a highlights list of human-readable facts Orion has learned — for example:

• "Revenue is your primary KPI"
• "CAC calculation excludes brand spend"
• "Always exclude cancelled orders from order counts"

These highlights are drawn directly from Orion's workspace memory — they are the actual rules Orion applies when answering your questions.

How to improve your score

The fastest ways to raise your score, in rough order of impact:

1. Certify your key reveals (5 pts each) — the highest per-item value. Certified reveals are the canonical definitions Orion trusts most.
2. Correct Orion when it is wrong (2 pts each) — every correction is captured and reused, and it makes future answers more accurate.
3. Record key metric definitions (3 pts each) — tell Orion "Remember that revenue excludes refunds" and it stores the rule.
4. Monitor more KPIs (2 pts each) — adding blessed KPIs both raises the score and enables silent-failure detection.
5. Keep asking questions (1 pt per 10) — usage compounds, and popular questions feed better autocomplete starters.

The breakdown panel includes a Teach Orion more button that opens the Orion bar pre-filled with Remember that… so you can add a fact in one step.

Admin view in Lumen

Luminaries get a fuller view at Lumen → Queries → Intelligence, with the complete component breakdown and the ability to edit workspace memory directly — add, correct, or remove the facts that drive the score.