Buying slices of a dataset

Slice-on-Demand lets you buy part of a dataset without buying the whole parent listing. A seller can publish a standing segment listing, or let you submit a row-match request for the keys you need. The feature is an opt-in beta and appears only on deployments where it is enabled. Production enablement remains blocked until the Slice Authorization rider is counsel-ratified and the production path is deployed.

This is a portal and REST workflow, not a private market-entry session. There are no buyer slice-request MCP tools in this demo.

1. Find a sliceable listing

Open Market → Search listings in the portal. A catalog-visible listing that is currently eligible for slicing opens a detail page with Buy only the rows you need.

The Market view loads a bounded catalog page, and its search box filters only the rows already loaded. It is not a global discovery index. A stable listing detail link can load one eligible listing directly by id, but it does not make a hidden or access-controlled listing public. When a listing is not currently eligible, its catalog projection is simply not sliceable; when the feature is off, the slice fields and portal controls are absent.

The detail page can offer either or both paths:

  • Buy segment purchases a standing seller-authored segment. It is buyable as an ordinary catalog listing after Amnetic has materialized the concrete, frozen child.
  • Request matching rows creates a row-match request from a query or key list. This path is useful for product names, SKUs, UPCs, account references, or any other keys the seller configured for matching.

2. Submit a row-match request

Paste one query per line, attach one non-empty CSV, XLSX, or TXT file, or do both. Pasted text and the file each have a 1 MiB limit. When both are present, the file is authoritative and the pasted text is ignored. The platform turns the accepted input into a canonical query list before it creates the job.

Request submission and matching are free; buying the resulting rows is not. The seller's policy limits the query count and the number of jobs your account can submit for that parent listing in a rolling 24-hour window. The server reports the limit that applies when it refuses a request.

Disclosure: Amnetic processes the submitted text or file into a canonical query list but does not deliberately store the raw uploaded file after interpretation. For every multi-column file, interpretation sends its column headers and up to five complete sampled data rows across all columns to the configured Claude model on AWS Bedrock. The canonical query list and job evidence are retained, and the seller can see that list and your requesting account identity through the seller-only activity and report surface. This beta publishes no query-retention or deletion control, and the seller-disclosure notice remains counsel-gated for production. A slice request is a disclosure-bound, named-buyer purchase order—not a confidential market-entry session. Submit only query material you are allowed to disclose to that seller.

After acceptance, the portal opens Slice requests for the new job.

3. Track coverage and review the offer

Slice requests shows a buyer-safe state: processing, offered, not_available, purchased, or expired. While processing, the fixed progress lines say only that Amnetic received the request, is matching it, or is preparing the result. Preparing your result can cover seller review or materialization; it does not expose internal worker or model output.

An offered result can show aggregate matched, near-miss, and unmatched totals. If the seller selected per-query disclosure, it can also show each submitted query with only its status. Neither view exposes seller row keys, candidate values, match confidence, row contents, or the rows themselves before purchase.

The displayed price is the server-frozen price of the child listing. That frozen child price is the binding quote: the portal does not reconstruct an authoritative quote from partial inputs, and a buyer does not send a price or account identity when purchasing it. Row-match offers expire; standing segment listings do not inherit that request expiry.

If the job becomes not_available, no purchase was made. The API reference owns the exact request limits, response shapes, and safe error taxonomy; see API reference.

4. Buy and download the slice

Buy the offered child with wallet credit or card. Card checkout remains pending until settlement and fulfillment confirm the child purchase. If settlement is still processing—or paid funds have been retained as wallet credit after fulfillment could not complete—the portal points to Billing & credits rather than claiming ownership.

After ownership is confirmed, download the slice as XLSX or Parquet. The receipt can list each matched query, its delivered key-column values, and the row's one-based position within the delivered slice. It never reveals the platform's internal row key or the row's position in the parent dataset. The receipt is reconstructed from hash-verified immutable evidence after purchase; it is not a separate stored artifact.

The MCP ownership_download tool still downloads only the canonical owned artifact. Explicit XLSX or Parquet selection is REST/portal-only. Turning the beta off stops new requests, offers, and slice purchases, but it does not revoke completed ownership; existing owners keep the normal, scan-gated download path.

Next

  • Selling slices — configure a dataset, build standing segments, and review buyer row matches.
  • API reference — exact buyer request, purchase, status, error, and download contracts.
  • Selling data — dataset intake and the broader seller surface.