Semantic Intelligence

An AI-built semantic layer for your databases.

Your database stores names. Semantic Intelligence adds meaning: what every table and column is for, the values a status column can hold, and the joins nobody declared. The Network Agent reads your schema and your real values, and Nova, the Query Streams AI built for database work, writes the layer for you, so Nova, Claude and ChatGPT query your database as it really is.

Get the free plan Team Pricing

Included on Business & Enterprise plans · Nova credits each month, then pay-as-you-go

Semantic Intelligence connector erp-prod · database sales coverage 92%

Captured

The Network Agent reads tables, views, columns, keys and indexes, on a schedule you choose and whenever you ask. This is all the database can tell you about itself.

Network Agent, on your network

ord_hdr · 2.4M rows

ohnoint cnoint stat_cdchar(1) ord_dtdatetime amt_totdecimal(12,2) rep_idint

cust_mstr · 38K rows

cnoint cust_nmvarchar(60) terrchar(3) cr_limdecimal(12,2) act_flchar(1)

0 foreign keys declared0 column commentsNo list of valid status codes

Understood

Nova describes every table and column, types each column, lists the values a code column really holds, and names the joins the Agent found. Nova is built for this: it knows every engine, dialect and connector Query Streams supports, from SQL Server and Oracle to file sets and Stripe. Anything you edit stays edited.

Written by Nova, you can override

ord_hdr — Sales order headers: one row per customer order, with its status, date, total and the sales rep who owns it. Fact table · Sales domain.

ohnoOrder numberidentifierprimary key

cnoCustomer who placed the orderjoins cust_mstr.cnomany-to-onechecked against the data

stat_cdOrder status ‘S’Shipped58% ‘O’Open31% ‘H’On hold7% ‘X’Cancelled4%

amt_totOrder totalmoney

rep_idSales rep who owns the orderjoins emp.emp_idchecked against the data

cr_limApproved credit limit, in the customer’s currencymoneyedited by your team

Joins found in the data 0 foreign keys declared · 4 candidates checked on your network

ord_hdr.cno cust_mstr.cnomany-to-onevalidatedEvery order’s customer number exists in the customer master.

ord_line.ohno ord_hdr.ohnomany-to-onevalidatedOrder lines roll up to their header.

cust_mstr.terr hr.territory.terr_cdmany-to-onevalidatedFound across two databases on the same server.

ord_hdr.rep_id cust_mstr.cno—rejectedThe numbers overlap by coincidence. Not a relationship.

Taught

Some knowledge isn’t in the schema: which table is retired, what counts as a test account, how a metric is defined. When someone corrects Nova in chat, Nova proposes a fact. It only takes effect once it has passed a read-only check against your data, or the person has confirmed a one-sentence restatement. Nova never just remembers.

Query Streams Verified Learning

Finance lead, in chat: “Don’t use ord_hist, it was retired. And leave test accounts out of revenue, they’re the ones with act_fl = 'T'.”

Nova: Checking ord_hist against ord_hdr on your network now. The test-account rule I’ll keep as: exclude customers with act_fl = 'T' from revenue and customer counts. Is that right?

Retired table verified against the data

ord_hist holds orders from 2014 to 2018 only; ord_hdr covers everything since. Nova stops reading ord_hist and keeps a one-line note for historical questions.

Checked on your network: the date ranges really are disjointRe-checked after every schema rescan

Business rule confirmed by the person who taught it

Test accounts have act_fl = 'T'. Exclude them from revenue and customer counts.

Author, evidence and every revision on recordEarns trust as answers that use it are accepted

Join reliability measured

ord_hdr.rep_id is blank on 11% of web orders, so a rep breakdown won’t add up to the total. Nova says so instead of hiding it.

Measurement re-run as your data changes

If answers that lean on a fact start getting corrected, the fact is suspended at once and re-checked. Facts nobody uses fade; facts people keep reaching for move up. Admins see every fact, its evidence and its author, and can revise, retire or block who may teach.

Asked

The same question, answered by an AI that only sees names, and by one that reads the layer. Only one of these runs.

Nova · Claude · ChatGPT

In plain English“How many orders did each customer have shipped last month?”

names onlyguessed

SELECT c.name, COUNT(*)
FROM   orders o
JOIN   customers c
  ON c.id = o.customer_id
WHERE  o.status = 'shipped'
GROUP BY c.name;

None of these tables or columns exist. Even with the right names, the status is a one-letter code, so the filter would match nothing.

Nova, with Semantic Intelligenceruns

SELECT c.cust_nm, COUNT(*) AS shipped
FROM   ord_hdr o
JOIN   cust_mstr c
  ON c.cno = o.cno
WHERE  o.stat_cd = 'S'
  AND  o.ord_dt >= '2026-09-01'
GROUP BY c.cust_nm;

The real table, the join found in the data, and 'S' from the column’s own value list.

Done on your network, by the Agent

Joins are found, then proven, on your network. ERP and legacy databases often declare no foreign keys at all. The Agent works out which columns could be keys, tests each candidate against your real values on its own machine, gives it a cardinality and a verdict, and rejects look-alikes that only overlap by chance. It can link tables across databases on the same server. Only the scores and a few example values leave, and the working data is deleted when the check ends.
It reads what your database already knows. Engine statistics stand in for table scans where they exist, and column comments you already wrote come along.
Built for production. Every query is a read the Agent builds itself, each one has a timeout, slow tables back off and retry, and discovery runs on a time budget.
SaaS and file sources too. Stripe, Shopify and HubSpot are profiled from the Agent’s local cache, so no vendor API limit is spent.

Sent to Nova to write the layer

Table and column names, types and keys
A few example values per column and a few sample rows, size-capped, long text truncated, binary replaced by a placeholder
Value lists for status and type columns, with their counts
Never a table copy. Exclude tables before a run, and purge Nova’s samples and descriptions by table, schema or database at any time.

Illustrative order-entry schema. Value lists come from the values actually in your columns, never from guesses.

Read by

Nova chat Nova in Query Builder

Claude over MCP

ChatGPT over MCP

A semantic layer you don’t have to build

Hand-built semantic layers are a modelling project. Semantic Intelligence starts from your data instead.

Hand-built layer (dbt, LookML)Semantic Intelligence

Who writes itYour data team, in YAML or LookMLNova, from your schema and real values. You edit what you want.

Getting startedA modelling project, usually on a warehousePick a database and switch it on

JoinsDeclared by handDiscovered and checked against the data

Where it worksThe modelled warehouseLive operational databases, file sets and API connectors, through the Agent

Keeping up with schema changesSomeone edits the modelFlags the database when it grows; the next run fills in only what’s new

Business rules and tribal knowledgeModelled by the data team, if anyone remembers toTaught in chat by the people who know, verified against the data or confirmed before Nova relies on it

Governed metric definitionsYes, that’s their strengthNot the goal. It describes your data and learns your rules; it doesn’t replace a metrics layer.

Already have a dbt or LookML layer? Keep it. Semantic Intelligence covers the databases it never reached.

Semantic Intelligence FAQ

What is a semantic layer?

A semantic layer sits between your raw tables and the people or tools asking questions. It says what each table and column means, how tables join, and which values are valid; some tools call the same thing a semantic model. Semantic Intelligence builds one automatically: the Network Agent reads your schema and real data, and Nova, the Query Streams AI built for database work, writes the layer. It adapts to every engine, dialect and connector we support, so AI tools write SQL that matches your database.

Will profiling slow down my production database?

It’s designed not to. Every query is read-only and built by the Agent, each one has a timeout, slow tables are retried later, and discovery runs on a time budget. Where your engine keeps statistics, the Agent reads those rather than scanning tables. You can also exclude large or sensitive tables before a run.

My database has no foreign keys. Does it still work?

Yes, and that is where it matters most. ERP and legacy schemas often have none. The Network Agent finds candidate keys, tests every candidate against the real values in both columns on your own network, and records the ones that hold up with a cardinality and a confidence score. Candidates whose numbers merely overlap by coincidence are rejected, and tables in different databases on the same server can be linked. Nova and the MCP tools only rely on the joins that passed.

What if Nova gets a description wrong?

Edit it. Descriptions and types you write are never overwritten by later runs. Nova also learns from corrections in chat, and each learned fact is checked against your data or confirmed before Nova relies on it. Administrators can review, revise or retire those facts.

Can Nova learn our business rules?

Yes. Query Streams Verified Learning is built into Nova. When someone corrects Nova or explains a rule in chat, Nova proposes a fact: a retired table, what a column really means, how reliable a join is, a business rule or a metric definition. Facts about the schema are checked with a read-only query on your network before they take effect; business rules take effect when the person confirms Nova’s one-sentence restatement. Every fact keeps its author, evidence and history. If answers that rely on a fact start getting corrected, it is suspended and re-checked, and measurements are re-run as your data changes. Administrators control who may teach and can revise or retire anything.

Which sources does it work with?

SQL Server, PostgreSQL, MySQL, MariaDB, Oracle, Snowflake, BigQuery, SQLite, Microsoft Access and DuckDB, plus file sets and the API connectors such as Stripe, Shopify, HubSpot and Google Analytics 4. It adapts to each engine’s SQL dialect.

Which plans include it, and who can run it?

Semantic Intelligence comes with Nova on Business and Enterprise plans. Runs count toward your plan’s data processing. Only administrators can switch it on, choose which databases and tables it covers, edit the results or purge them. Everyone who uses Nova or MCP benefits.

Nova AI MCP Server Query Builder Reports & Data Alerts Database Integrations

Let Nova learn your database.

Connect a database, switch on Semantic Intelligence, and ask your first question in plain English.

Get the free plan Team Pricing

Included on Business & Enterprise plans · Nova credits each month, then pay-as-you-go