Glossary

Product objects

Data Language Model (DLM): a foundation model that understands tables natively, from raw cell values, and returns understanding, relationships, missing values, and predictions in one pass. Schema models are DLMs.

Base: a Schema model generation, named schema-{n} (s{n} in report names). Runs and new endpoints use the latest base unless you name one. Models.

Quick run: POST /v2/run. A stateless pass over the tables in the request that returns the full bundle and persists nothing. Its prediction is labeled tier: "in_context". Quick run.

Endpoint: your persistent object: a base, pinned dataset snapshots, a configuration (target, task, thresholds), a library of system prompts, and a held-out report, served at a stable URL. Endpoints.

Serve: POST /v2/serve/:id. Live inference over the request tables, scored against the endpoint’s pinned data, returning the bundle. Needs the serve scope.

Refresh: POST /v2/endpoints/:id/refresh. Re-syncs the pinned snapshot, scores held-out on a fresh split, and mints a new report. The only operation that makes an endpoint stop reflecting removed rows.

Upgrade: POST /v2/endpoints/:id/upgrade. Re-runs the same pinned data on a newer base and mints a new report.

System prompt: a named instruction fragment (sp_...) on an endpoint that steers an LLM. Executed in the platform’s Chat; exposed on the endpoint object so your own LLM can apply it.

Job: an async unit of work (job_...) for endpoint creation, refresh, upgrade, synthetic generation, and large or batch runs. Poll GET /v2/jobs/:id. Jobs.

Batch: a service tier elected per job with "options": { "processing": "batch" }. Runs with a completion window and bills the batch rate ($1.50 / Mcell).

Scope: what an API key may do: read, run, serve, manage, delete. Authentication.

Data

Connection and dataset: a connection (conn_...) is where data comes from; a dataset (ds_...) is an ingested or registered table. Both live under /v2/data.

Snapshot: an immutable copy of a dataset minted by each ingest or sync, id dsv_.... Endpoints pin snapshots and held-out reports stamp them.

Pinned: a dataset state. Pinned data serves repeat calls at the cached rate indefinitely and counts toward plan storage. Endpoint creation pins its datasets automatically. Data.

Outputs

Bundle: the one output shape every run, creation, and serve returns: sector, column profile, cross-table map, missing-value imputation, and predictions, each with its own confidence. Outputs.

Pass: one inference over a set of tables. Every pass emits the full bundle; there is one rate per pass, with no per-output charge.

In-context: how every prediction is made: from the labeled rows present (the request’s, or the endpoint’s pinned data). Nothing trains.

Target and task: the column to predict and the prediction type (classification, regression, anomaly). Both default to auto; the response states what was chosen and why.

Sector: the industry or domain a table belongs to, returned as top1 and top5 with a confidence each. Detection is sector-agnostic: it understands the values, not the column names.

Column profile: the per-column understanding: role, role_confidence, type, missing_pct, pii.

Cross-table map: how two or more tables relate without shared keys: column alignments, entity matches, and a unified schema, each with a confidence.

Column alignment: a pair of columns from two tables that carry the same attribute, with its confidence and a verbatim example from each side.

Entity match: a pair of rows from two tables that describe the same real-world entity, with the attributes it was matched on.

Unified rows: the cross-table map materialized into merged records, one per matched entity, with provenance per field and the raw variants of shared values.

Missing-value imputation: filled values for empty cells, numeric and categorical, with a row-level confidence. Original cells are never overwritten.

Trust

Held-out: the score on rows whose target values the model did not see when predicting them, measured by an in-context split at creation, refresh, and upgrade, stated with its split, seed, baseline, and snapshot. Held-out reports.

Report: the immutable trust artifact minted by each create, refresh, and upgrade. Id r_..., display name {endpoint}.{base}.{op}.{YYYY-MM-DD}. Reports.

Metering

Cell: one value, one row by one column. The metering unit. 1 Mcell = 1,000,000 cells. Missing cells count like any other cell. Usage and billing.

Fresh and cached: fresh cells are understood for the first time and bill the fresh rate ($3.00 / Mcell); cached cells are unchanged pinned data re-used on a later pass and bill the cached rate ($0.30 / Mcell).