Semantic Intelligence

Give your AI context.
Teach Nova your business.

Semantic Intelligence describes your tables, explains their codes and checks their relationships, giving Nova and AI tools through MCP context for your real database. In Nova, ordinary conversation can also become verified business knowledge for later questions. No SQL or table names needed.

Semantic context: Nova + MCPConversational learning: Nova

Comes with Nova, the AI analyst built into Query Streams, on Business and Enterprise plans.

Your schema, with meaning
ar_invtabledescribed Customer invoices. One row per invoice, with the customer it went to, when it was issued, what is still owed and its status. Finance domain.
inv_noint
Invoice numberidentifierprimary key
cust_idint
Customer the invoice was sent tojoins cust_mstr.cust_idmany-to-onechecked against the data
joins
cust_mstrCustomers. One row per customer account.
cust_idCustomer number, primary key
cust_nmCustomer name
terrSales territory code
inv_dtdatetime
Date the invoice was issueddate
amt_duedecimal(12,2)
Amount still owed on the invoicemoney
stat_cdchar(1)
Invoice status
‘O’Open ‘P’Paid ‘D’Past due ‘X’Cancelled
Four values, read from the column itself

Names become descriptions Nova can use.

Illustrative schema and meanings. On your database, the agent reads the actual structure and values; Nova generates descriptions you can review and edit. See how you can teach Nova, too.

Read on your network. Written by Nova.

Two parts do the work. The Network Agent, a small program you run inside your own network, reads the database where it lives. Nova, the AI analyst built into Query Streams, turns what it sends back into plain words.

  • The agent reads the schema in placeTables, views, columns, keys and indexes, on a schedule you choose and whenever you ask. Column comments you already wrote come along.
  • Joins are tested on your own machineThe agent works out which columns could be keys and checks each candidate against your real values. Only the scores and a few example values leave; the working data is deleted when the check ends.
  • Built for production databasesEvery 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. Where your engine keeps statistics, the agent reads those instead of scanning tables.
  • Nova writes it. You can overrule it.Nova describes every table and column, types each column and names the joins the agent proved. Anything you edit stays edited; later runs never overwrite it.
From database evidence to useful context
Database and Network Agent inside a cutaway network enclosure, connected to a sheet of table and column descriptions
Read beside your dataCheck the relationshipsWrite useful descriptions

The Network Agent reads schema and selected values inside your network. Database credentials stay there.

Nova + your AI through MCP
Shared context for tables, codes and relationships.

Three things it works out.

The parts of a database that nobody wrote down, and that an AI tool otherwise has to guess.

What each table and column means

A sentence for the table, a line for every column, and a type for each: identifier, date, money, code. Edit any of it and your words stay.

The values a code column really holds

Status and type columns are read from the data, with the count of each value, so a filter uses the one-letter code the database actually stores.

The joins nobody declared

Older and ERP databases often declare no foreign keys at all. Every candidate join gets a cardinality and a verdict, look-alikes that only overlap by chance are rejected, and tables in two databases on the same server can be linked.

The same question, two answers.

Names alone leave room for the wrong assumptions. Here, the descriptions, checked join and real status codes give Nova the context to write the query on the right. Follow the highlighted evidence into the SQL.

Follow the evidence into the query

Asked in plain English“How much is still owed by each customer on past-due invoices?”

Names onlyguessed
SELECT c.name, SUM(i.amount)
FROM   invoices i
JOIN   customers c
  ON c.id = i.customer_id
WHERE  i.status = 'overdue'
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.
With Semantic Intelligenceruns
SELECT c.cust_nm, SUM(i.amt_due) AS owed
FROM   ar_inv i
JOIN   cust_mstr c
  ON c.cust_id = i.cust_id
WHERE  i.stat_cd = 'D'
GROUP BY c.cust_nm
The real tables, the join proved against the data, and ‘D’ from the column’s own value list.

Table and column descriptions point to ar_inv, cust_mstr and amt_due.

Who reads the layer: Nova, in chat and in the Query Builder, and the AI tools you already use, like Claude, ChatGPT or Cursor, through the MCP server. The conversational learning system is exclusive to Nova.

Conversational learning, only in Nova

Just tell Nova.
It learns the meaning.

Explain your business, correct an answer or describe how your team works. You don’t need SQL, table names or column names. With teaching enabled, Nova can turn ordinary conversation into useful semantic knowledge for later questions.

Semantic Intelligence supplies database context to Nova and AI tools through MCP. This conversational learning system is a Nova feature. Your organization controls who may teach; administrators can review the evidence, revise knowledge or retire it.

Meet Nova, your AI analyst

Learning has its own verification system.

Nova checks facts it can prove with read-only database checks. Business rules and metric meanings are confirmed with you in plain English. A proposal has to pass the required verification before it becomes active knowledge. If later corrections cast doubt on a fact, it can be suspended and re-checked.

A conversation becomes useful knowledge
Finance lead

Our test customers are only for practice. Leave them out of what customers owe us.

Nova

Exclude test customers when calculating amounts owed.

Is that what you mean?

Learning verificationBusiness meaning confirmed
  • Nova uses the database context to identify test customers.
  • You confirm the business meaning in ordinary conversation.
  • The confirmed rule is saved with its author and evidence.

Saved for later: business rule

Evidence, author and revision history on record
Later, in NovaWhat is our past-due balance?
Using verified business rule
Illustrative result after using the learned knowledge
CustomerPast-due balance
Castellan Foods28,900.00
Northgate Clinics17,320.80
46,220.80 owed

Nova applies the saved rule. The sample test account’s 8,640.00 is left out without you having to repeat the instruction.

See the query Nova handled
SELECT c.cust_nm, SUM(i.amt_due) AS owed
FROM ar_inv i
JOIN cust_mstr c ON c.cust_id = i.cust_id
WHERE i.stat_cd = 'D'
  AND c.act_fl <> 'T'
GROUP BY c.cust_nm

Independent illustrative scenarios and sample data; no query is sent and no knowledge is saved. Your explicit descriptions take priority over generated descriptions. Verified learning keeps its evidence and history for review.

What leaves your network. What never does.

Nova needs enough to describe a table, and no more. You can exclude tables before a run, and purge Nova’s samples and descriptions by table, schema or database at any time.

Sent to Nova to write the layer

  • Names, types and keysTable and column names, their types and the keys the database declares.
  • A few example values per columnAnd a few sample rows, size-capped, long text truncated, binary replaced by a placeholder.
  • Value lists for status and type columnsWith the count of each value.
  • Join scoresThe verdict and a few example values from each candidate the agent checked.

Never

  • A copy of a tableThe agent reads where the data lives and sends a summary.
  • The working data from a join checkIt is deleted on your machine when the check ends.
  • Your database loginIt stays with the agent inside your network.
  • Anything from a table you excludedChoose the databases and tables before the run.

Works with the databases you have.

Live operational databases, not only a modelled warehouse. Nova knows each engine’s dialect, so the layer fits the database it describes.

Database engines and hosted connections
SQL Server
PostgreSQL
MySQL
Oracle
Snowflake
BigQuery

File sets and the API connectors such as Stripe, Shopify, HubSpot and Google Analytics 4 are covered too. SaaS sources are profiled from the agent’s local cache, so no vendor API limit is spent.

Shopify
Stripe
HubSpot
On Business and Enterprise, with NovaFrom Business Core at $50 a month
  • Switch it on per database. Pick a database and turn it on; there is no modelling project first.
  • Administrators run it. Only administrators can switch it on, choose which databases and tables it covers, edit the results or purge them.
  • Everyone benefits. Anyone who uses Nova or connects an AI tool through the MCP server gets the better answers.
  • Counted like other work. Runs count toward your plan’s monthly data processing. Nova comes with credits each month, then pay-as-you-go.

Compare every plan

Questions people ask about Semantic Intelligence

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 for you: the Network Agent reads your schema and real data, and Nova writes the layer, so AI tools write SQL that matches your database.

Will profiling slow down my production database?

It is 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. The 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. Nova and the AI tools connected through the MCP server 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.

We already have a hand-built semantic layer. Do we need this?

Keep it. A hand-built layer is a modelling project, usually on a warehouse. Semantic Intelligence starts from your data instead and covers the live databases your layer never reached. It flags a database when it grows, and the next run fills in only what is new. It is not a governed metrics layer: it describes your data and learns your rules; it does not replace metric definitions.

Give Nova the context behind your data.

Connect a database on the free plan. Semantic Intelligence comes with Nova on Business and Enterprise plans, and switches on per database when you are ready.