Skip to content
P
Back to Guides

AI: ingestion, enrichment & BYOK

Extract products from documents, enrich copy and translations, and bring your own model keys with per-function routing.

PIE uses AI in three places: ingestion (turn a document into a product), enrichment (improve product copy), and onboarding field mapping. All of them run through one provider adapter, and every organisation controls which provider and model they use.

Providers and BYOK

The AI adapter supports a built-in free tier plus bring-your-own-key (BYOK) for OpenAI-compatible and Anthropic providers. Configure providers from your organisation's AI settings: producers (pip_admin) and buyers (pic_admin) manage their own org's settings; a platform admin can act cross-tenant. AI settings are not grant-scoped — they never resolve across organisations.

With BYOK you can set per-function model overrides, so ingestion, enrichment, field mapping, and translation can each use a different model. The free tier is model-locked.

Ingestion

Upload a document (PDF, XLSX, DOCX) or paste a URL. The pipeline preprocesses the input, then the provider extracts structured JSON, which stops at a mandatory review step. You can edit any extracted field — including fixing an invalid GTIN check digit inline — before approving. Approval creates or updates products (new, existing, or merged). Review is never skippable, so nothing is imported unseen.

Enrichment

Run enrichment on a product to generate a description, SEO copy, quality-agent gap suggestions, or translations into your other enabled locales. Each suggestion lands in a review queue where you accept or reject it; the review page polls and updates without a manual refresh. Accepting a translation writes through to the per-locale attribute columns and re-scores compliance, so localized data stays consistent with its evaluation.

Bulk enrichment lets you select products by filter (category, status, query, tags) and enqueue them. The selection is resolved server-side under row-level security and hard-capped, and an over-cap id list returns a clear error rather than silently truncating.

Rate limits and resume

AI calls respect per-minute, per-day, and token-based rate limits. When a provider throttles, jobs queue and resume automatically rather than failing outright.

Related