Product
Sembric runs one governed workflow: discover what you already have, classify it, certify exactly what agents may use, then enforce that on every call. Here is each step and the rules it enforces.
1 · Discover
An owner connects a BigQuery dataset, Snowflake schema, Trino catalog and schema, Databricks catalog and schema, PostgreSQL schema or MySQL database, and tests the customer identity each Sembric role uses before saving.
2 · Classify and steward
Columns with common personal-data names start sensitive. That heuristic is a starting point, so review by the asset’s assigned data steward is mandatory before anything built on it can be certified.
3 · Define
Both carry a business definition, caveats, concepts, role grants and at least two distinct expected-answer cases. Every saved definition is versioned with a field-level history.
Aggregate
One measure — sum, count, count_distinct, avg, min or max — over approved dimensions. Sum, average, minimum and maximum need a numeric measure so their checks are verifiable.
A reviewed minimum group size is enforced in the source query with HAVING COUNT(*) >= n for previews, certification, the API and MCP. It withholds small groups but is not complementary suppression, so keep person-level dimensions coarse.
Projection
Approved, non-sensitive columns with optional fixed row restrictions. Restrictions are equality or value lists, never open ranges, and are always conjoined with what the caller asks for.
Certification also proves that each granted source identity can actually read the approved projection — a row count alone cannot prove column access.
4 · Certify
A builder requests review from a named, independent owner or engineer. The request closes itself when the surface is certified or its definition changes.
Data product
Net revenue by region
| Check under role | Expected | Result |
|---|---|---|
| Analyst · Q1 total | 711,290.00 | pass |
| Analyst · North only | 849,000.00 | pass |
| Viewer · Q1 total | 412,880.00 | pass |
| Viewer · North only | 849,000.00 | pass |
Reviewer is neither the owner nor the last editor · fingerprint unchanged through the final check
5 · Serve
Agents and applications query certified surfaces through MCP or the REST API. Sembric builds the source query; callers supply only these inputs.
| Rule | What Sembric enforces |
|---|---|
| Fields | Only declared dimensions or approved columns; a product query may group by at most three of up to 12 declared dimensions. |
| Filters | Equality, a value list of at most 50, or a range on numeric, date and time fields — always bound as typed parameters. |
| Time grains | day, week, month, quarter or year on declared date or timestamp dimensions; weeks start Monday and instants bucket in UTC. |
| Limits | 1 to 100 rows. A query matching more is refused with narrowing guidance instead of being truncated. |
| Result shape | Provider rows must have exactly the approved columns; an extra or missing field fails closed. |
| Response size | Serialized rows are capped at 1 MB by default (configurable up to 4 MB) before release. |
| Fixed restrictions | A view’s equality or value-list restrictions are always conjoined with caller filters and cannot be replaced. |
After a source query returns, Sembric rechecks membership, role, the calling session or key, and the certificate before releasing rows.
Context for agents
Concepts declared on certified surfaces form a shared ontology. Owners and engineers connect them with typed relationships, and certification refuses a conflicting definition for an existing concept.
Workspace
Membership and role are re-read on every request, so changes take effect immediately for people and their agents.
| Role | Can |
|---|---|
| Owner | Connects sources and maps role identities, assigns stewards, invites and removes members, clears sensitive marks, revokes any key. |
| Engineer | Browses the catalog, rediscovers sources, marks columns sensitive, builds and certifies surfaces, monitors agents. |
| Analyst | Sees and queries the certified products and views granted to the analyst role. |
| Viewer | Sees and queries the certified products and views granted to the viewer role. |
FAQ
A data product is one aggregate measure — sum, count, distinct count, average, minimum or maximum — over approved dimensions. A governed view projects approved columns with optional fixed row restrictions. Both carry business context, role grants and expected-answer checks.
An owner or engineer who is neither the surface’s current owner nor its last definition editor. They run two distinct expected-answer checks under every granted role’s source identity before certifying.
Thirty days by default. In the last seven days an independent reviewer can renew it by re-running every check. An expired certificate is withdrawn, recorded and never renewed; the surface returns to draft.
A product can set a reviewed minimum group size, enforced in the source query with HAVING COUNT(*) >= n. It withholds small groups but is not complementary suppression, so keep person-level dimensions coarse.
Tell us which warehouse you run and what your agents need to answer. We will help you connect a source, certify a first surface and query it over MCP.