Product catalog and manufacturer questions
The product-research page now has a separate catalog panel. Research dossiers and their quoted evidence, question documents, review receipts and downgrade protections keep their existing behavior.
A catalog identity is manufacturer + model + explicit variant. Matching normalizes Unicode NFC, case and whitespace; punctuation and variants remain distinct. Discovery titles and aliases cannot merge products. Concurrent imports serialize per project and create immutable evidence versions under one canonical identity. CRM companies remain manufacturer/business records, never product records.
Each specification retains its exact quoted assertion, source URL, excerpt SHA-256, capture time, format, declared attribution, model/variant match, typed number/unit when recognized, confidence and missing/contradicted state. Supported means the excerpt states an exact identity, not that the manufacturer or claim was verified. kg/g and m/cm/mm normalize; arbitrary values stay text. Manufacturer/community attributions are declared claims. Conflicting claims remain visible even after a human confirmation. New evidence never changes the confirmed claim. Freshness is recalculated on reads (30 days; future captures are unverified). Version readback uses ?version=N; each version records changed fields.
The catalog form performs at most one search and three fetches through the existing pinned public-web extraction interface. The API allows up to eight fetches. Search snippets are discovery only. The maintained web adapter currently retains a paragraph/metadata brief capped at 1,500 characters; only explicitly labeled specifications in that excerpt are parsed, and absent or unsupported layouts remain unknown. Uploaded PDF bytes use the maintained isolated PDF extraction worker, with a 400,000-byte input cap, 12,000-character excerpt cap and its existing parser deadline/page/memory limits. Public PDF downloading remains unavailable; /pdf accepts already acquired document bytes. No new crawler or model tools are granted by evidence. Every source is untrusted data.
Missing, contradicted, unverified or stale fields produce a versioned UNSENT manufacturer request, linked to the product version and field names. Review uses an interactive human session and a current request version; changed products require a fresh request. The composer contract has no recipient and send:false. Actual mail sending is OPEN until the private composer and verified recipient setup are integrated. A returned evidence record binds request, conversation reference, product and fields, appends a product version and changes the request to needs_human_review. A conversation reference is unverified input, not proof of the sender. Supported returned excerpts do not automatically confirm a specification. Requests show a seven-day follow-up due date and overdue state; no reminder or email is dispatched.
All catalog routes sit under /v1/orgs/{orgId}/projects/{projectId}/product-research/catalog, inherit the artifact flag/auth/no-store boundary, and enforce current org/project membership. Reads require viewer, writes tester, confirmation/review an interactive session. Mutations recheck credential and creator standing before commit. A collection reserves an immutable operation UUID before I/O. Stop, revoked authority, deadline or an ambiguous operation prevents later writes. Repeating a completed operation returns its stored result; stopped/UNKNOWN operations never replay external extraction. A fresh operation is an explicit new intent, not an automatic retry.
Integration status (2026-10-20): official migration13510 and the five-table catalog schema are deployed. Live acceptance on ad811c76344cd40c74739d856da613a69a0cf6e1 exercised a synthetic QA import, actual browser confirmation, reload and historical-version readback. Scoped API/MCP acceptance on 8d7b04735aa37ff6fa9ce6994e9d272fdf2d19f6 used an ordinary read-only key for canonical REST reads and the three curated tools product_catalog_list, product_catalog_read and product_catalog_operation; a valid import was denied403, and revocation was followed by denied401 readback. These catalog paths remain present in the subsequently deployed 5c5db530137776cbd80ee6b55b5267d6f370e888 release. This proves the scoped catalog journeys, not complete QAC or manufacturer-outreach execution.
Curated MCP access requires the normal read/mcp/artifacts scopes and current creator, org, project and role ceilings. The six mutation tools additionally require write scope; an API key cannot human-confirm a specification, review a manufacturer question or send mail. The typed daily action delegates scheduling and approval/egress checks to the existing deterministic workflow executor; complete scheduled catalog-to-manufacturer execution remains OPEN. No independent scheduler was added. Verified-recipient mail delivery, authenticated reply ingestion and the full provider/workflow journey remain OPEN. The recorded catalog acceptance used synthetic QA records and no live search, model, email send or manufacturer contact.
- API reference: Catalog routes in OpenAPI.
- MCP tools: Curated catalog tools and permission boundaries.
- Runbook: Self-host release, verification and rollback operations.
Implementation relied on the current PostgreSQL transaction-lock documentation and Zod strict object schemas. Source-specific SQL, HTTP, MCP, PDF and browser behavior is exercised against locked local dependencies by the owning catalog tests.