GFS Platform

Cloudflare

Workers, D1, KV, R2, Pages, DNS, Access — inventory, limits, AI query roadmap

Products in use

4 active, 3 partial

Planned

AI query layer

Critical gap

Access on Pages

Account ID

9d209df5…

Plan

Workers Paid

$5/mo base

Why this page exists

Cloudflare is the load-bearing layer of the platform — Worker (API), D1 (data), KV (cache), R2 (files), Pages (UI), DNS (routing), and soon Workers AI + Vectorize for the natural-language query layer. This page consolidates every product, its current state, where it's referenced in source, common operations, and what's planned next. Decisions about Cloudflare are tracked in ADRs; live health is on admin-dashboard.

Inventory 11 products

Per-product card with status, instance, binding, source/config, common ops, and limits.

Workers

Active

gfs-platform

Single public API for the platform. 20+ GET endpoints over the D1 mirror. CORS allowlist, Bearer auth, security headers.

Detail

Binding: —

Source: src/index.ts

Config: wrangler.jsonc

Limits: 100,000 requests/day on free tier; CPU 50ms/request; sub-requests 50/request (Paid: 1000)

Common ops:

wrangler deploy
wrangler tail
wrangler dev

Endpoint detail →

D1

Active — 185K rows

gfs-netsuite

SQLite read-mirror of NetSuite. 18 tables, 152 GL accounts, 102K transactions across 20 types, full entity backfill.

Detail

Binding: DB

Source: schema.sql, sql/*.sql (45 loaders)

Config: wrangler.jsonc d1_databases

Limits: 10 GB max per DB (currently ~150 MB), 100K reads/day free, point-in-time recovery on Paid

Common ops:

wrangler d1 execute gfs-netsuite --remote --command "SELECT…"
wrangler d1 backup create gfs-netsuite
wrangler d1 export gfs-netsuite

Tables, columns, FK, queries →

KV

Provisioned, not used

CACHE (e880e40b…)

Short-TTL cache layer. Bound but no endpoints read/write yet. Planned uses: cache /api/ar/aging, /api/kpis responses; rate-limit counters for AI query layer; AI response cache.

Detail

Binding: CACHE

Source: src/index.ts (env.CACHE)

Config: wrangler.jsonc kv_namespaces

Limits: 100K reads/day free, 1K writes/day, 25 MB value max

Common ops:

wrangler kv:key put --binding=CACHE "k" "v"
wrangler kv:key list --binding=CACHE

Binding declaration →

R2

Active (light use)

gfs-files

Binary file storage. File Upload RESTlet (NS script 2512) writes here via SuiteAttach. Planned uses: scheduled D1 → SQL export for DR, archived PDF templates, AI audit log cold storage.

Detail

Binding: STORAGE

Source: src/index.ts (env.STORAGE)

Config: wrangler.jsonc r2_buckets

Limits: 10 GB free storage, 1M Class A ops/month, 10M Class B/month (egress is free)

Common ops:

wrangler r2 object put gfs-files/path file.bin
wrangler r2 object list gfs-files

Pages (legacy)

Deprecated — pre-consolidation

gfs-system-guide.pages.dev

Original platform site. Kept available until migration is fully validated. Will be deleted after greenlight.

Detail

Binding: —

Source: Frozen at pre-consolidation state

Config: —

Limits: Same as Pages (above)

Common ops:

dash.cloudflare.com → Workers & Pages → gfs-system-guide → Delete

Legacy URL →

DNS / Zone

Active

ai-globalfoodsolutions.co

Authoritative DNS for ai-globalfoodsolutions.co (owned). Corporate domain globalfoodsolutions.co also owned. 6 records: CNAME (api, www), MX, TXT (SPF + DMARC), NS.

Detail

Binding: —

Source: wrangler.jsonc routes

Config: dash.cloudflare.com → ai-globalfoodsolutions.co → DNS

Limits: DNS records: 3,500/zone soft limit

Common ops:

dash.cloudflare.com → DNS → add/edit records

Record table →

Access (Zero Trust)

Tier 1 GAP — required before sharing URL

(not configured)

Auth in front of gfs-netsuite.pages.dev. Currently anyone with the URL can read GAPS, NS reference, runbook. Tier 1 in GAPS_TO_CLOSE.

Detail

Binding: —

Source: —

Config: dash.cloudflare.com → Zero Trust → Access → Applications

Limits: 50 free seats on the Free Zero Trust plan

Common ops:

Create Application → Self-hosted → gfs-netsuite.pages.dev → Email OTP policy → Save

Tier 1 GAP entry →

AI Gateway

Planned — AI query layer

(planned)

Proxy + cache + logging layer in front of all LLM calls. Lets us swap models (native Llama ↔ Claude/GPT) without changing app code.

Detail

Binding: —

Source: —

Config: dash.cloudflare.com → AI → AI Gateway → Create

Limits: Free tier covers caching, observability, and most use; per-request pricing for advanced routing

Common ops:

Create gateway
Point Worker LLM calls at gateway URL

ADR-011 →

Vectorize

Planned — AI query layer

(planned — gfs-corpus)

Vector store for the embedded corpus (D1 entities + manifest + research docs + SuiteQL library + schema). 768-dim (BGE-base). Source of retrieval for AI query.

Detail

Binding: VECTORIZE (planned)

Source: —

Config: wrangler.jsonc vectorize (planned)

Limits: Free tier: 30M queried dimensions/month, 5M stored vector-dimensions

Common ops:

wrangler vectorize create gfs-corpus --dimensions=768 --metric=cosine
wrangler vectorize insert gfs-corpus --file=embeddings.ndjson

ADR-011 →

Workers AI

Planned — AI query layer

(planned)

Native inference at the edge. Embeddings: @cf/baai/bge-base-en-v1.5. LLM: @cf/meta/llama-3.3-70b-instruct (fallback @cf/meta/llama-3.1-8b-instruct).

Detail

Binding: AI (planned)

Source: —

Config: wrangler.jsonc ai (planned)

Limits: Pay-per-neuron; ~$0.59/M in + $0.79/M out for Llama 3.3 70B (verify before launch)

Common ops:

env.AI.run('@cf/meta/llama-3.3-70b-instruct', {messages})
env.AI.run('@cf/baai/bge-base-en-v1.5', {text})

ADR-011 →

AI query layer — phased build

Natural-language query over D1 + the manifest + research docs. Decision recorded in ADR-011 (native Cloudflare stack: Workers AI + Vectorize + AI Gateway). Phased so each step ships independently.

Phase	Deliverable	Status	Detail
Phase 0	Cloudflare surface + ADR	Done	This page + ADR-011 in decisions log.
Phase 1	Vectorize index seeded	Planned	Create gfs-corpus, embed manifest (3,486 items) + 8 research docs + 29 SuiteQL queries via Workers AI bge-base. Stored vectors ~10K, well under free tier.
Phase 2	Schema + entity chunks	Planned	Add 18 table descriptions + FK notes + ~2,500 entity rows (customers/vendors/items/employees/gl) as additional chunks. Aggregate-only on transactions (monthly summaries, not row-level).
Phase 3	Query endpoint v1	Planned	POST /api/ai/query — retrieve+answer mode only (no SQL gen). Cite source row/doc. Bearer auth. Per-token rate limit via KV.
Phase 4	SQL generation mode	Planned	LLM emits parameterized SELECT → validator (allowlist tables, no DDL/DML) → D1 execute → LLM summarizes. Test corpus: 20 known admin questions with expected SQL.
Phase 5	/ai-query.html UI	Planned	Chat surface. Citations as links. 'See generated SQL' expandable. Per-role: admin sees SQL by default, reader sees answer only.
Phase 6	Audit + status endpoints	Planned	/api/ai/audit, /api/ai/status, /api/ai/sources. KV 30d hot, R2 cold.
Phase 7	Reindex automation	Planned	Worker scheduled() cron — incremental reindex of D1 rows modified since last run.

Default decisions (you can override before any phase starts)

LLM: Llama 3.3 70B Instruct (native Workers AI) for v1. AI Gateway lets us proxy to Claude/GPT later without app changes.
Embedding model: BGE-base-en-v1.5 (768d, native). Swap to text-embedding-3-small via AI Gateway if recall is poor.
Transactions in corpus: aggregate-only (monthly summaries per customer/item/account). Row-level is too big and answers little.
Show generated SQL: yes for admin role, no for reader. Builds trust without overwhelming non-technical users.
Reindex cadence: nightly Worker cron for v1. Live updates (on sync) in v2.
Audit retention: KV 30 days hot, R2 cold indefinitely.

Endpoints (planned)

Endpoint	Auth	Purpose
POST /api/ai/query	Bearer	Single-turn natural-language query → answer + citations + optional SQL
POST /api/ai/index	Bearer (admin)	Force reindex (one source or all)
GET /api/ai/status	Bearer	Last index date, query count 24h, cache hit rate
GET /api/ai/sources	Bearer	List indexed sources with chunk counts
GET /api/ai/audit	Bearer	Recent queries (last N), redacted

Customer pricing pipeline — phased build

Derive customer price lists from invoice history → review in platform → push back to NS. Decision recorded in ADR-012. Per-customer lists (no shared Price Levels), NetSuite stays the daily editing surface, platform handles migration + audit + AI review.

Phase	Deliverable	Status	Detail
Phase 1	Derive job + diff endpoint	Done	Schema migration applied (customer_pricing_proposed, customer_pricing_history, pricing_anomalies). Derive produced 661 rows across 103 customers / 219 items (445 median_last_3, 216 last_paid). Worker endpoints live: GET /api/pricing/derive, /api/pricing/diff, /api/pricing/customer/:id, /api/pricing/audit/summary. Buckets: 434 listed+recent (drift), 830 listed+stale (deprecate), 227 not-listed+recent (orphans — biggest gap). 90 open SOs across 33 customers flagged for cutover handling.
Phase 2	/pricing.html audit view	Planned	Read-only surface. Per-customer view shows Listed+Recent / Listed+Stale / Not-Listed+Recent buckets. Sortable, searchable, exportable CSV. No write to NS yet. ~1 day.
Phase 3	Approval flow	Planned	Admin marks rows approved / rejected / edited in /pricing.html. Writes to customer_pricing_proposed.status. Still no NS push. ~1 day.
Phase 4	Push to NS via SuiteAPI RESTlet	Planned	Worker bulk-writes approved rows to NS Customer → Item Prices sublist. Per-row success/failure reporting. Pilot on 3 customers, validate next-day SOs, then expand. ~2 days.
Phase 5	Effective-date history	Planned	customer_pricing_history table populated on every push (price, effective_from, effective_to, source, change_reason, changed_by). GET /api/pricing/history/:customer/:item. ~half day.
Phase 6	AI review pass	Planned	Workers AI flags anomalies (drift > threshold, outside-peer-band, suspicious one-offs), suggests smoothing for promo-polluted rows, surfaces stale customers due for review. Renders in /pricing.html. ~1 day.
Phase 7	Quote letter generation	Planned	From approved price list + customer context, AI drafts a PDF/HTML quote letter. Stored in R2 with customer + effective date. Optional email via Email Routing. ~1 day.

Locked defaults

Derive rule: median(last 3 invoices) when ≥3 exist in 18-month window, fall back to last_paid otherwise.
Unsold items: left off the customer's price list; populated at next quote/SO.
System of record: NetSuite for live orders. D1 holds proposed + history. Platform reviews + pushes.
Exclusions: credit memos, returns, negative quantities. Bid pricing (custbody1 Bid Reference) tagged separately so AI can distinguish.

New D1 tables

customer_pricing_proposed — (customer, item, derived_price, source, source_invoice_id, derived_at, status, approved_by, approved_at, notes)
customer_pricing_history — (customer, item, price, effective_from, effective_to, source, change_reason, changed_by, changed_at) — populated on every push to NS
pricing_anomalies — (customer, item, anomaly_type, severity, detected_at, resolved) — AI's running output

New endpoints

Endpoint	Auth	Purpose
POST /api/pricing/derive	Bearer (admin)	Rebuild customer_pricing_proposed from invoice_lines
GET /api/pricing/diff	Bearer	Three-bucket audit across all customers (Listed+Recent, Listed+Stale, Not-Listed+Recent)
GET /api/pricing/customer/:id	Bearer	Current + proposed for one customer
POST /api/pricing/approve	Bearer (admin)	Mark proposed rows approved / rejected / edited
POST /api/pricing/push-to-ns	Bearer (admin)	Worker → SuiteAPI RESTlet bulk write
GET /api/pricing/history/:customer/:item	Bearer	Full effective-date history

Pre-cutover audit (Phase 1 output)

Before any write to NS, the diff endpoint surfaces:

Listed + Recently paid: stable rows. Drift check (paid price within tolerance of listed price?).
Listed + Stale: customer hasn't bought this item recently. Candidate to deprecate from list.
Not-Listed + Recently paid: customer is buying it but it's not on their price list. Negotiated at SO time. Biggest gap to close.
Open SOs in flight: separate query, surfaces SOs that haven't invoiced yet so cutover doesn't break in-progress orders.

AI's role

Anomaly detection — "last paid 30% below median, investigate" (Workers AI + customer_pricing_history)
Smoothing suggestions — propose median over last_paid where the spread is large
Peer comparison — "12 customers in this program pay $5.45 for item X, customer Y pays $5.20"
Quote letter generation — draft PDF/HTML quote from approved price list (Phase 7)
Drift narrative — explain why customer X's price for item Y changed N times this year (reads customer_pricing_history)
Never auto-approves. Human review is mandatory before any push to NS.

Other Cloudflare roadmap

Non-AI items that depend on Cloudflare products. Sequencing per the GAPS tracker tiers.

Item	Tier	Detail
Worker scheduled() handler for sync	Tier 2 GAP	Move sync.sh into the Worker so it calls SuiteAPI directly. Eliminates the single-Mac dependency. ~4 hr.
R2 archival of D1	Tier 2	Daily Worker cron: wrangler d1 export + put to R2. Rolling retention 30/90/365. ~2 hr.
KV caching wrapper	Tier 3	Wrap /api/kpis, /api/ar/aging, /api/items/performance with KV (TTL 5 min). ~3 hr.
Email Routing / Email Workers	Optional	If dunning notices move out of NS or admin alerts (sync failed, gap created) need to go to email. ~2 hr.
Browser Rendering API	Optional	If Advanced PDF/HTML limits get hit, render templates via Browser Rendering. ~half a day to wire.
Durable Objects	Defer	Only if real-time multi-user state shows up (live dashboard collaboration, in-app chat). Not currently needed.

Usage & limits

Approximate current usage vs ceiling. Refresh from the Cloudflare dashboard for live values.

Product	Metric	Current	Ceiling	Headroom
Workers	Requests/day	low (manual)	unlimited (Paid)	∞
D1	Storage	~150 MB	10 GB / DB	98.5%
D1	Reads/day	low (sync only)	5M (Paid)	~100%
D1	Writes/day	low (sync only)	100K (Paid)	~100%
KV	Reads/day	0	10M (Paid)	100%
R2	Storage	light	10 GB free, then $0.015/GB-mo	large
Pages	Builds/month	~5	500 free	99%
Pages	Concurrent builds	1	5 free	large

If you cross 80% on any line, the Worker should start logging a warning and the admin dashboard should show a yellow pill. Tracked as a sub-item under GAPS.

Cost model

Current monthly burn (Workers Paid plan $5 base). All other usage is within free tiers.

Line item	Unit cost	Current usage	Monthly $
Workers Paid plan	$5/mo base	—	$5.00
D1 reads / writes / storage	free up to limits	well under	$0.00
KV ops	free up to limits	0	$0.00
R2 storage	$0.015/GB-mo after 10 GB	under 10 GB	$0.00
Pages builds + bandwidth	free	under limits	$0.00
Workers AI (when AI query ships)	per-neuron, ~$0.001–$0.005/query	0 today	~$5–30 est at 100 queries/day
Vectorize	$0.04/M queried dims; storage tiny	0 today	~$1–5 est
AI Gateway	free tier covers cache + observability	0 today	$0.00
Access (Zero Trust)	free up to 50 seats	0 (not configured)	$0.00

Wrangler quick-ref

# Auth + identity
wrangler login
wrangler whoami

# Worker
wrangler deploy
wrangler tail
wrangler dev                                # local

# D1
wrangler d1 execute gfs-netsuite --remote --command="SELECT 1"
wrangler d1 backup create gfs-netsuite
wrangler d1 export gfs-netsuite --output=backup.sql

# KV
wrangler kv:key put --binding=CACHE "key" "value"
wrangler kv:key get --binding=CACHE "key"
wrangler kv:key list --binding=CACHE

# R2
wrangler r2 object put gfs-files/path/file.bin --file=local.bin
wrangler r2 object list gfs-files

# Pages
wrangler pages deploy . --project-name gfs-netsuite
wrangler pages deployment list --project-name gfs-netsuite

# Secrets
wrangler secret put API_KEY
wrangler secret list

# Vectorize (planned)
wrangler vectorize create gfs-corpus --dimensions=768 --metric=cosine
wrangler vectorize insert gfs-corpus --file=embeddings.ndjson

Cross-links

Data model — D1 schema, ER diagram, example queries, SuiteQL admin library
Infrastructure — Worker endpoints (per-endpoint detail), DNS records, bindings, secrets
Admin dashboard — live health of every Cloudflare service
ADR-011 — AI stack decision
GAPS — Tier 1 (Access on Pages), Tier 2 (Worker-side sync, D1 backups)
Runbook → incident response — sync stopped, D1 quota, Pages down

Cloudflare

Why this page exists

Inventory 11 products

Workers

D1

KV

R2

Pages

Pages (legacy)

DNS / Zone

Access (Zero Trust)

AI Gateway

Vectorize

Workers AI

AI query layer — phased build

Default decisions (you can override before any phase starts)

Endpoints (planned)

Customer pricing pipeline — phased build

Locked defaults

New D1 tables

New endpoints

Pre-cutover audit (Phase 1 output)

AI's role

Other Cloudflare roadmap

Usage & limits

Cost model

Wrangler quick-ref

Cross-links

Related surfaces

Infrastructure

Data model

Decision log

Admin dashboard

Runbook