Dataset Intelligence Quickstart

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

For installation and initial provider connection, begin with the canonical install, connect, and verify guide.

Start in ordinary language

Tell AIWG what you have and what you want:

Use ./exports/support.ndjson for searchable support history, and make every
answer traceable to its source. Keep everything local.

The `dataset-intelligence` router creates an intake, proposes a bounded source check and redacted preview, and returns a recommendation with rationale, assumptions, materialized artifacts, privacy/locality/network implications, and safe defaults.

Review before writes

Register and assess the source, then create an immutable plan:

Use AIWG to complete this documented outcome: Register and assess the source, then create an immutable plan
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

Review the exact plan digest, schemas, estimated reads and writes, artifact classes, capability fallbacks, retention, and authorization references. Policy-sensitive work pauses for approval scoped to that digest.

After approval, ingest using the exact digest and a caller-selected idempotency key:

Use AIWG to complete this documented outcome: After approval, ingest using the exact digest and a caller-selected idempotency key
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

The service reports run, receipt, rejection, cancellation, and checkpoint references. It does not execute a connector or index data itself.

Resume an incremental source

Ask:

Sync the support-history dataset from its last committed checkpoint, show me
what changed, and do not treat missing records as deletions without review.

The incremental flow reassesses the source, builds a new immutable plan, and requires review of changed-source evidence, tombstone policy, prior checkpoint, and expected writes. Cancellation preserves the last committed checkpoint.

Explain and verify

Ask “where did this result come from?” to request a bounded lineage query. Evidence includes identities, revisions, schema and run bindings, assertion basis, method, confidence, privacy, and locators.

Ask “is this dataset and index current?” to request verification. The result distinguishes verified, degraded, unverifiable, and failed states. Search success alone is never verification.

Use AIWG to complete this documented outcome: Ask “is this dataset and index current?” to request verification. The result distinguishes verified, degraded, unverifiable, and failed states. Search success alone is never verification.
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

Export or retire

An export uses a versioned standards profile and reports exact mapping coverage and loss. Descriptor-only profiles are not executable capabilities.

Retirement first creates an the agent-owned dataset operation plan with complete affected-artifact enumeration, canonical/derived distinctions, holds, tombstones, bulk threshold, and rollback limits. Nothing is deleted by an agent or skill. The shared service performs only the approved operation and returns reconciliation evidence.

Local JSONL/CSV adapters are stable and the checked-in local orchestration, offline, provenance, and standards cells are qualified. Pre-stable migration, Fortemi Core parity, and live Fortemi Server persistence remain pending. See the task guide support matrix and consult the agent-owned dataset operation for the installed action surface. The aggregate conformance receipt is not stable-eligible while those three cells remain pending.