Cognitive Kernel

Natural LanguageIn-Process ModelSemantic Search

Roadmap Planned · not yet usable · aspirational in-process AI kernel; verify availability before relying on it

What it is

The Cognitive Kernel is a small language model that runs inside the database process. Its job is to turn an intent — “show me the sales trend in Karnataka for the last three years” — into a query the engine runs, then write a plain-English answer over the result. Because it runs in-process, your schema and rows don’t leave the database to answer a question.

Every response is structured: the narrative answer, the raw data behind it, and a logic trace explaining the steps taken — so a reviewer can audit any answer rather than trust the model blindly.

flowchart LR
  classDef u fill:#1f2640,stroke:#9ca3af,color:#e8eaf0
  classDef k fill:#1f2640,stroke:#7c5cff,color:#e8eaf0
  classDef e fill:#1f2640,stroke:#00d4ff,color:#e8eaf0
  classDef o fill:#1f2640,stroke:#4ade80,color:#4ade80

  USR[Natural-language intent]:::u
  BRAIN[In-process model]:::k
  PLAN[Query over your tables]:::k
  ENG[Engine executes]:::e
  MMC[Structured answer<br/>narrative + data + trace]:::o

  USR --> BRAIN --> PLAN --> ENG --> BRAIN --> MMC

Why it matters

Natural-language querying usually means shipping your schema and rows to an external model API — a data-egress and trust problem. The Cognitive Kernel runs the model inside the database process, so questions are answered without your data ever leaving, and every answer comes with the query it ran and a step-by-step logic trace you can audit.

Data never leaves — the model runs in-process; there’s no external LLM endpoint receiving your rows.
Auditable, not a black box — each answer ships the narrative, the raw data, and the reasoning steps, so a reviewer can check the work.
Grounded in your schema — questions resolve to real queries over real tables, not a guess.

How you use it

The ask built-in is the SQL surface for natural-language queries. It returns a table with three columns — narrative, data, and trace.

-- Ask a question in plain English.
SELECT narrative, trace
FROM ask('What were the top 5 SKUs by revenue in Karnataka last quarter?');

Plain SQL still works too — semantic similarity over text is available as a function:

SELECT id, description,
       semantic_similarity(description, 'durable hiking shoes') AS sim
FROM   products
ORDER BY sim DESC
LIMIT 10;

For deliberately broad, exploratory questions you can opt into a larger external model that sees only anonymized aggregates — never raw rows:

SELECT *
FROM ask(
  'Compare our regional revenue trends to industry benchmarks and suggest where to invest.',
  use_bridge => TRUE
);

Key concepts

Concept	What it means
`ask(...)`	The SQL surface for natural-language queries; returns narrative, data, and trace.
Logic trace	The auditable chain of steps taken to produce the answer, queryable as a column.
External bridge	An opt-in escalation to a larger model that receives only anonymized aggregates.

Notes & caveats

Semantic NLP →

How text columns become AI-ready.

Multimodal Fabric →

The same model runtime, for media.

Legacy Transmuter →

How the kernel translates legacy procedures.

Compute →

The in-database execution model.