

How Orion uses context: the right context, at the right time, for the right reason
Have 10 people in your org ask an AI tool "What was revenue last week?" and you might get 10 different numbers, depending on which context each one happened to grab. Orion is built so that does not happen.
How general-purpose LLMs handle context
ChatGPT, Claude, and similar tools let you set up projects and attach context to them: files, instructions, notes. That context sits in one pile. The model decides on each turn what looks relevant, with no rules about which piece should answer which kind of question.
That works for writing and research. It breaks down for analytics. Take "What was revenue last week?" The pile might hold a finance definition in one doc, a bookings export from another system, and a dashboard screenshot built on a third source. Each one is a plausible way to answer. Which one the model uses depends on what it pulls in on that turn, so the same question can come back with different numbers on different days, each delivered with the same confidence.
Orion works differently. Context in Orion has structure, and the product carries instructions that tell Orion when to pull in a given piece of context, based on the question and the user's answers along the way. Ask Orion about revenue last week and it goes to the governed revenue definition, every time. The goal is the right context, at the right time, for the right reason, so the answer is grounded in your business and not invented.
The difference is not how much context each tool can hold. It is whether anything decides which context answers which question.

Three kinds of context, three jobs
Orion reads context from three places. Each one is built for a different kind of question, and Orion knows which is which.
Every question passes through the same rules, lands on the source built for it, and comes back with a citation.
The metric catalog: what a number means
Every metric in Orion carries its definition, the logic that computes it, and the data sources behind it. Revenue, churn, pipeline, CAC: each one is defined once and used the same way in every analysis, by every user, in every channel.
When Orion uses it: any time a question touches a governed KPI. Orion starts from the catalog definition instead of writing its own version of the math. If your team already defines metrics in LookML or dbt, that logic is the starting point. Bring your semantic layer, and Orion uses it to write better SQL.
Why it matters: the most common analytics error is not bad SQL. It is good SQL for the wrong definition. The catalog removes that choice from the model.
The Knowledge Base: how the business works
The Knowledge Base holds the things a new analyst learns in their first month: SOPs, business rules, how fiscal periods work, which customers are excluded from which reports, how the team prefers to analyze a problem.
Teams write pages directly in Orion or import them from Notion, Confluence, and GitHub, as a one-time import or a daily sync. The docs you already maintain become context Orion can use, without anyone copying them over by hand.
When Orion uses it: when the answer depends on how your business operates, not just on the data. Pages are enabled per project, so a Finance project reads Finance rules and a Marketing project reads Marketing rules. A Marketing attribution rule never leaks into a Finance close analysis.
Why it matters: Orion cites the pages it used, so anyone reading the answer can see which rule shaped it and fix the page if the rule is out of date.
State pages: what is true right now
State pages are structured context: tables that Orion and your team both read and write. Think of a tracker of open investigations, the list of campaigns live this quarter, current account owners, or a log of known data issues.
When Orion uses it: when the answer depends on an exact fact that changes week to week. Who owns this account? Is this campaign still running? Did someone already investigate this drop? Orion reads the answer from a row, not from a paragraph it has to interpret.
Why it matters: prose is good for explaining a rule. It is a poor home for facts that need to be exact and current. State pages also record who changed what, so the context Orion relies on has a history, and the work Orion does can be written back for the next person, or the next agent, to pick up.
Knowing when matters more than knowing more
Giving a model more context does not make it smarter. Giving it rules about when to use which context does. Orion's product carries those rules, so its reasoning follows a pattern a good analyst would recognize.
The right time. Orion does not load every piece of context up front. It pulls context in as the question and the user's answers make it relevant, the way an analyst opens the definitions doc only when a number is in question. Less noise in front of the model means fewer chances to grab the wrong thing.
The right reasoning. Context shapes how Orion works a problem, not just what it knows. A Knowledge Base page on how your team investigates a revenue drop tells Orion which cuts to check first, so its reasoning follows your playbook instead of a generic one.
The right means. Each kind of context is stored in the shape that fits it: definitions as governed logic, rules as written pages, live facts as tables. Orion reads each in that native form, so nothing gets lost turning a formula into a paragraph or a table into a summary. Structure is also where models work best: AI took off in programming because code is structured, and State pages give your team's custom requirements that same footing.
In practice. Ask Orion "What was revenue last week, and why did it drop?" and it pulls the revenue definition from the catalog, applies the Finance team's fiscal week rules from the Knowledge Base, and checks a State page for known data issues before hunting for a cause. Ask again next week, or from Ops, and only the data changes.
No hallucinations first
Analytics hallucinations rarely look dramatic. They look like a reasonable number that happens to be wrong, and each one traces back to a missing piece of context. Churn calculated with a guessed formula is a missing metric definition. An explanation built on a policy nobody wrote down is a missing business rule. "The campaign is live" when it ended two weeks ago is a stale fact. The metric catalog, the Knowledge Base, and State pages each close one of those gaps.
When Orion lacks the context a question needs, it says so or asks instead of guessing. That is what lets a business user act on an answer without waiting for an analyst to double-check it.
Bring the context you already have
Most teams already wrote this context down. It lives in LookML, dbt, Notion, Confluence, and the tracker someone updates every Monday. Orion turns it into structured context with rules for when to use it, so every answer starts from what your business actually means. And because that context is shared, every person and every agent on the team works from the same definitions. That is what makes analytics multiplayer.
Already defining metrics in LookML or dbt? Orion uses that logic as context from the first question. See how the Knowledge Base works, or book a demo to see it on your own data.

