Integration · Data & infra
Cloudflare Workers AI, D1 and Vectorize: the stack this site runs on
This website is a static export on Cloudflare Pages, and its assistant, contact form and analytics run as Pages Functions on the Workers runtime. We build the same way for clients who want AI at the edge with no servers to patch.
Cloudflare
Data & infra
Access
Authentication and least-privilege scopes
Bindings on the Pages project or Worker instead of keys in code, plus scoped API tokens for deploys
- deploy token: Pages edit only
- D1 + Vectorize edit: provisioning only
- bindings: per environment
- secrets: encrypted variables
Code reaches D1, Vectorize, KV and Workers AI through bindings, so the repository holds no database passwords or AI keys.
Objects
What we read and what we write
Pages project (static assets)
Read: YesWrite: No
HTML, CSS and media served from Cloudflare’s network.
Pages Function / Worker
Read: YesWrite: Yes
API routes such as chat, contact and events on the Workers runtime.
D1 database
Read: YesWrite: Yes
SQLite-compatible store for leads, sessions and the full-text index.
Vectorize index
Read: YesWrite: Yes
Embeddings for semantic search next to the D1 full-text search.
Workers AI binding
Read: YesWrite: Yes
Model and embedding calls with no API key in code.
KV namespace
Read: YesWrite: Yes
Small, fast lookups such as rate-limit counters.
Data flow
System of record and write-back
The static site and its API ship as one deployment. D1 is the system of record for leads and sessions, and Vectorize holds embeddings rebuilt from the site’s content.
- Static pages are served from Cloudflare’s network.
- API routes run as Functions on the Workers runtime, close to the visitor.
- Functions read and write D1, query Vectorize and call Workers AI through bindings.
- Anything that needs a person, such as a lead or a handoff, is stored in D1 and sent by email.
- Visitor browser
- Contact form
- D1
- Vectorize
- Workers AI
The steps below trace one question to this website’s own assistant, as its code handles it.
Worked example
Cloudflare in a build
Step 01
This page, served from the edge
miniailabs.in is a static Next.js export on Cloudflare Pages, so the page arrived as ready-made HTML with no server render.
Step 02
The assistant posts to the chat Function
A Pages Function checks the origin and the request size, then applies a per-visitor rate limit kept in KV.
Step 03
Redact and route
Card numbers, IDs and keys are masked. Greetings, off-topic questions and handoff requests get a fixed reply without a model call.
Step 04
Retrieve
An in-memory index of the site content is searched, and D1 full-text and Vectorize results join it by rank when those bindings exist.
Step 05
Answer from numbered sources
The model is called through Workers AI or another configured provider, with the retrieved passages as its only sources.
Step 06
Check and cite
An output check blocks banned claims, unlisted prices and unknown links, then adds citations; low confidence offers a person.
Step 07
Store the turn, if allowed
With a database bound, the turn is kept under a hashed session id with an expiry; the visitor can delete it.
Sample workflows
What teams build with Cloudflare
Pattern 01
A cited assistant on your own site
Hybrid search over your pages in D1 and Vectorize, a model call through Workers AI and an output check before anything is shown.
Pattern 02
Lead capture with an audit trail
Form posts are validated at the edge, stored in D1 with a reference number and sent to your team by email.
Pattern 03
Edge APIs for small tools
Calculators and planners run as Functions beside the static site, with no separate server.
Failure handling
What breaks, and what the build does about it
A binding is missing in an environment
Code checks each binding and degrades: retrieval falls back to the bundled index, storage is skipped, and the response still works.
A burst of requests from one visitor
Per-visitor rate limits in KV return a clear error before any model call is made.
The model provider fails
The next configured provider is tried; if none answers, the visitor gets a short message and upstream error details are never echoed.
Fit · Cloudflare
When it is the wrong choice
- Long-running jobs or heavy CPU work per request. Use a queue or a container.
- A relational workload that outgrows one D1 database or needs Postgres features. Use Postgres, for example Supabase.
- A specific large model that Workers AI does not offer. Call that provider from the Worker instead.
Alternatives worth pricing
Postgres with pgvector and row-level security for multi-tenant apps.
When the model you need is not on Workers AI; the Worker calls it directly.
A container platform
Long-running or CPU-heavy work that does not fit the Workers runtime.
Where it is used
Related use cases, services and guides
Limits
Limits and common errors
A Pages site can hold 20,000 files on the Free plan and 100,000 on paid plans, at up to 25 MiB per asset; Pages Functions requests count towards your Workers quota.
Source: Pages limits (opens in a new tab) · Cloudflare Docs · accessed
Workers get 10 ms of CPU per request on Free; on Paid the default is 30 s, raisable to 5 minutes. Each isolate has 128 MB of memory.
Source: Workers limits (opens in a new tab) · Cloudflare Docs · accessed
A D1 database can reach 10 GB on Workers Paid (500 MB on Free), with up to 1,000 queries per Worker invocation (50 on Free).
Source: D1 limits (opens in a new tab) · Cloudflare Docs · accessed
D1 supports the FTS5 module for full-text search.
Source: D1 SQL statements (opens in a new tab) · Cloudflare Docs · accessed
Vectorize allows up to 1,536 dimensions per vector and 20,000,000 vectors per index; topK is 50 when values or metadata are returned and 100 without.
Source: Vectorize limits (opens in a new tab) · Cloudflare Docs · accessed
Workers AI is billed in Neurons, a measure of GPU compute, with a free daily allocation on every account.
Source: Workers AI pricing (opens in a new tab) · Cloudflare Docs · accessed
Security
Security notes
- Bindings replace secrets in code; the few real secrets, such as the email API key, are encrypted environment variables.
- Every D1 query binds its parameters, so visitor input is never concatenated into SQL.
- Chat turns are stored only under a hashed session id with an expiry, and visitors can delete them.
FAQ
Frequently asked questions
Does this website really run on Cloudflare?
Yes. It is a static export served by Cloudflare Pages, and its API routes, including the assistant and the contact form, run as Pages Functions. Any response from miniailabs.in carries Cloudflare’s headers.
Can Workers AI replace an external model API?
For many tasks, yes: classification, summaries, embeddings and grounded answers from retrieved text. When a task needs a model it does not host, the Worker calls that provider instead, behind the same interface.
Is D1 suitable for production data?
It is a managed SQLite-compatible database with a size limit per database, which suits leads, sessions, settings and search indexes. Large relational workloads belong in Postgres.
How does hybrid search work on Cloudflare?
Full-text search in D1 finds exact words, Vectorize finds passages with similar meaning, and the two ranked lists are merged so a passage that scores well on either rises. Only the top merged passages reach the model.
What does a Cloudflare build cost to run?
It depends on traffic. Static hosting is inexpensive, while Functions, D1, Vectorize and Workers AI are billed by use. The proposal estimates the running cost from your expected traffic using Cloudflare’s current pricing pages.
Next step
Have a workflow in mind? Let’s scope it together.
Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.
Free 30-minute call · no commitment