Metabase
Metabase made querying accessible to non-analysts. A conversational layer takes the remaining step for people who will never open it.
Metabase made querying accessible to non-analysts. A conversational layer takes the remaining step for people who will never open it.
What Metabase holds, and why that matters here
Metabase is widely used because it lowered the barrier to asking data questions — saved questions, dashboards and a query builder that does not require SQL. It works well for people who will engage with a tool.
A substantial population will not. They will ask a colleague. For them, the saved questions already defined in Metabase are exactly the right foundation for a conversational answer.
- Saved questions — the queries already defined and validated.
- Dashboards — assembled views.
- Models — curated data structures.
- Collections — how questions are organised.
- Permissions — who may see which data.
Questions Elbi can answer from Metabase
What is the number for this?
A saved question retrieved conversationally.
How many this month?
A period figure from an existing query.
What is the trend?
Direction from an existing time series.
Which is the largest?
A ranked answer from a defined question.
When was this last updated?
Data currency.
Is there a question for this?
Pointing to existing work rather than duplicating it.
How Elbi works with Metabase
- Staff ask, not the publicMetabase usually sits over internal databases, so this is a staff-facing connection. People sign in through your own identity provider before asking anything.
- Saved questions are the vocabularyThe assistant is given a named set of saved questions and models from specific collections, each described in Arabic and English with the phrasings people use. A question is answered by running one of those, with the period or filter the person mentioned.
- An API key with a narrow groupMetabase API keys belong to a permission group and act with that group's rights. The key used here sits in a group that can view the chosen collections and has no permission to create new queries, native SQL included.
- Audiences are kept apartBecause a key acts as its group rather than as the individual asking, audiences that must see different data get separate groups and keys. Where row-level restrictions matter, that separation is designed before launch, not assumed.
- Freshness is statedMetabase queries the source database directly unless caching or model persistence is switched on. The answer says when the result was produced, and if it came from a cache, says so.
- No question, no answerIf nothing in the catalogue fits, the assistant says so and points to the data team, or to the dashboard nearest the need. It does not improvise a query.
Scope and control in Metabase
| Direction of access | Runs saved questions and reads models in the collections granted. Nothing is created, edited, archived or shared in Metabase. |
|---|---|
| Query rights | The API key's group has no query-builder or native SQL permission, so ad hoc queries are blocked by Metabase's own permissions, not merely discouraged by instruction. |
| Who can see what | Collection and data permissions of the key's group set the outer limit. Different audiences use different groups, mapped to people through your sign-in. |
| Credentials | A Metabase API key created by your administrators. You can regenerate or delete it from Metabase's admin settings, and the connection stops when you do. |
| Where data goes | Results are used to answer and then discarded with the conversation: message content is destroyed within 24 hours, usually when the conversation ends, and no external AI service sees it. |
| Load on your database | Saved questions run against the database Metabase points at. If that is production, heavy questions stay out of the catalogue, or run against a replica or warehouse instead. |
What this replaces
The natural fit is that Metabase already contains curated questions. Those are named, validated and owned — precisely the shape a conversational layer should expose, and precisely what stops an assistant from inventing its own analysis.
Building on saved questions rather than on raw queries also keeps answers consistent with what the organisation already reports, which avoids the second-set-of-numbers problem entirely.
The population served is people who need a figure occasionally and will not learn a tool for it — which in most organisations is considerably larger than the population that will.
Where Metabase questions come from
Startups and scale-ups
Where Metabase is the analytics layer and adoption is uneven.
Operations teams
Occasional figures without opening a tool.
Management
Numbers on request.
Any Metabase deployment
Reaching people who will never open it.
Getting value from Metabase saved questions
Look at which saved questions and dashboards people already open most, and which they ask a colleague to open for them. The overlap — a figure that is looked up often, by people who do not use Metabase themselves — is the first catalogue.
Expect to tidy before connecting. Collections that have grown for years usually hold several versions of the same question with slightly different filters. Choosing the one that is right, naming it clearly and archiving the rest helps everyone who uses Metabase, not only the assistant.
A useful test of the catalogue is whether a manager new to the organisation could use it. If a saved question's name only makes sense to the analyst who built it, the assistant will struggle to match it to a plain request, and so will everyone else; renaming it is usually the cheapest improvement available.
Then add "is there a question for this?". Pointing someone to an existing dashboard is cheap, keeps the data team's work in use, and is often more useful than a single number.
What connecting Metabase involves
- Pick the collectionsChoose the collections whose questions and models are maintained and trusted, and name an owner for each.
- Describe each questionGive every saved question in scope an Arabic and English description, the filters it accepts, and the phrasings that should lead to it.
- Create the group and keyAn administrator creates a permission group with view access to those collections and no query-creation rights, then an API key assigned to that group.
- Settle row-level needsDecide whether different people may see different rows. If they may, create a group and key per audience; if not, confirm that every answer is fit for everyone who can ask.
- Check the database loadTime the questions in the catalogue and move anything slow to a replica or warehouse before the assistant starts running them on demand.
- Start with one teamRelease to the team that interrupts the data function most, and add questions as the logs show what else is asked.
Personal data behind Metabase
Metabase is often connected straight to application databases, which means raw tables with names, phone numbers, email addresses and national ID numbers sit one click from the questions built on them. Under the Saudi Personal Data Protection Law, an assistant has no reason to reach any of that to answer "how many orders this week?".
The group behind the API key should therefore see curated models and the collections that use them, not the underlying tables. Questions that return individual records belong outside the catalogue unless there is a specific, agreed reason for them.
Metabase's own options, and where an assistant fits
Metabase already has good ways to push figures to people: dashboard subscriptions by email or Slack, alerts when a question's result crosses a goal or returns rows, and embedding charts in other applications. Its query builder is also approachable enough that many staff do answer their own questions.
An assistant reaches the rest — people who will not open Metabase, ask in Arabic, or need the figure in the middle of a conversation about something else. Built on saved questions, its answers match the dashboards exactly.
The version that fails is the one given broad data access and asked to write its own queries. It produces plausible numbers that nobody has checked, which is precisely the problem a curated Metabase is meant to prevent.
What data teams see with Metabase connected
Common questions
It uses your saved questions, which keeps answers consistent with what you already report.
Yes, bounded by what the person asking may see.
It says so. Defining the question is analytical work that belongs with someone who understands the data.
No. The API key is yours, created in your Metabase and limited to the collections its group can view. It cannot write queries, and deleting it ends the connection.
Only the network path, which is agreed during deployment. The same API key and group permissions apply, and nothing needs to be moved.
Rarely. Metabase usually reports on internal data, so the connection is set up for signed-in staff. Customer-facing figures are better served from the system that owns them.
Connect Metabase to Elbi
Your saved questions, answered in a sentence for signed-in staff, in Arabic and English.