How PartsIQ works
A durable, auditable pipeline. The language model reads and writes natural language; the database and code decide every fact โ compatibility, stock, quantity and approval.
The request pipeline
Parse the enquiry; extract items, language, urgency.
Resolve the machine + configuration from text/serial/photo.
Retrieve ONLY catalogue-compatible parts (lexical + BGE-M3).
Rank candidates; bounded, read-only specialist tools.
Allocate across warehouses; show shortfalls honestly.
Model writes prose slots; code inserts every fact.
Seven named code checks decide pass/fail.
A human approves the current revision before publish.
Services
gateway
The only public surface. Verifies the Supabase token, enforces roles/assignment, calls use cases. Never runs heavy work inline.
search
Preserved legacy identification + compatible-set filtering + fulfilment planning, plus BGE-M3 semantic retrieval.
orchestrator
The workflow + the deterministic verification gate + approval/revision transactions. Pass/fail decided in code.
worker
Nameplate OCR, approved-PDF export, catalogue import/publish โ behind leases with fencing tokens.
apps/web
Customer chat + demo, staff review workspace; bilingual (en/ar), theming, responsive.
Deterministic verification gate
Before a recommendation can reach a human, it must pass seven named, pure code checks. An LLM is never asked whether verification passed; a failed check surfaces to the reviewer, never silently treated as success.