From supplier documents to versioned quotes

LLM-first extraction, deterministic quoting. Claude Sonnet 5 (via Vercel AI Gateway, with a direct-Anthropic fallback) reads uploaded PDFs, Excel files, images, CSVs, and DXF geometry and emits structured JSON against a Zod schema. The quoting engine downstream is pure – no LLM in the price math. This separation is the load-bearing decision: extraction can be improved model-by-model without touching the rate math, and the rate math is auditable.

Streaming structured output. The /api/extract-case route uses AI SDK's streamText with output: Output.object({ schema }). The client-side useObject hook deserializes the partial JSON as tokens arrive, so the estimator sees rows populate live rather than a spinner for 20 seconds.

Prompt caching everywhere. System prompt + schema description cached as text blocks; PDF files cached as file blocks under Anthropic's 5-minute TTL. Cold-path latency on a 51k-token payload is 10–20s; cached retries drop to 3–5s. The Vercel function has a 300s max duration (Fluid Compute) to accommodate the cold path without failing.

Versioned rate cards with a single "current" per workspace. The rate_cards table has a unique index on (workspace) WHERE is_current – exactly one active card per workspace, enforced by the database rather than the app. Concurrent saves race on the constraint and get a 409, which prevents "someone saved while I was editing" without any wall-clock timestamp drift.

Rate-key normalization with dual-standard fallback. Extraction emits material names verbatim from supplier docs ("A516 GR.70", "S420MC / H420LA"). A normalization pass strips casing, whitespace, and punctuation noise; a lookup then tries exact match first, then normalized, then dual-standard splitting for A / B composite names against cards that only list A. Without this, near-duplicates flooded the "missing rate" queue.

Alternatives considered and rejected:

• Rules-based extractor – too brittle against supplier document variance; would have required a per-supplier parser per PDF revision.
• LLM in the price math – closes off auditability. The client has to be able to say "here is the rate we charged you, here is the supplier source, here is the date."
• Client-side extraction – PDF files can be large, browser memory is not the extractor's friend, and the AI Gateway routing has to live server-side.

• Email-driven intake. Inbound RFQ emails hit a Resend inbound address; a Svix webhook (signature-verified) writes the email + attachments to Supabase, with attachments in a private Storage bucket. The dedup key is Resend's messageId on a UNIQUE index – idempotent by construction.
• Four-format ingestion. PDFs and images encode as base64 file blocks; spreadsheets (Excel, CSV) are flattened server-side to CSV text before hitting the model; DXF geometry parses to a structured payload for the quoting engine.
• Immutable snapshots on quote issue. The issued_quotes table stores case_snapshot and rates_snapshot alongside the current live rate card. A quote issued in June answers the question "what did we quote you against?" three months later without archaeology.
• Bilingual by construction. All UI strings are Loc = { nb: string; en: string } objects consumed by useLang(); inbox emails detect BCP-47 locale and render the correct envelope copy. Norwegian is the default (hydration-safe for SSR).
• RLS deny-all on all four tables. No authenticated or anon policies. The service-role client behind auth middleware is the only ingress. The database enforces the boundary; the app enforces which rows a session may touch.

• Cost per extraction is logged. Every LLM call goes through recordLlmGeneration() to Langfuse with input/output tokens and cost. Per-extraction spend is visible in the fleet dashboard.
• Formula-injection guard on any XLSX export. Cells beginning with =, +, @, or – are sanitized. Estimators paste extracted rates into other Excel workbooks; a mis-typed material name should not exfiltrate a workbook.
• Deployment protection. noindex header, custom subdomain, Vercel deployment protection. The tool is not searchable and not linkable.

1. Structured LLM extraction at production quality – Zod-schema constrained output, streaming to the client, aggressive prompt caching for cost + latency.
2. Postgres schema design for versioned data – unique partial indexes for "current" flags, immutable snapshots on quote issue, event-sourced status transitions.
3. Multi-format ingestion pipeline – PDF, image, Excel, CSV, DXF, email attachments, all normalized to structured JSON before touching the price engine.
4. Security-first webhook handling – Svix signature verification, private Storage buckets, RLS deny-all with service-role gating.
5. Bilingual product from day one – nb/en toggle in the UI, BCP-47 detection on inbound emails, Norwegian-default rendering.

1. Put the LLM in the extraction path, keep it out of the money math. Auditability is a feature, and once it's gone it's expensive to restore.
2. Enforce single-current invariants in the database, not the app. A unique partial index is one line and catches a whole class of race.
3. Ship the paid-API cost log before the paid-API feature. Every LLM call in production is a rounding-error decision that adds up; instrument it before you can be surprised by it.