Semantic Image Search

semantic-image-search

Vercel's Semantic Image Search — a page of images searched by what they show rather than by their file names: a vision model titles and describes each image, the description is embedded, and a search is embedded and ranked against it by cosine similarity in pgvector, beside a plain text match — running on Cloudflare Workers + Static Assets through @opennextjs/cloudflare, its images in an R2 bucket served by the Worker, its searches cached in Workers KV, both models Workers AI's through the Worker's AI binding, and its vectors in a Supabase project's Postgres reached through Cloudflare Hyperdrive — all provisioned by Terraform through Orun, the schema applied by Orun's db-migrate, the way the rag-chatbot baseline does it, and packaged as an Orun baseline: one command takes a new product from nothing to live on stage and prod, in its own repository, under its own name — in ONE landing.

The template's code is the code. What the port changes is what the Workers runtime and a fresh unattended deployment force. Upstream indexed by hand: three scripts (upload to Vercel Blob, generate-metadata with GPT-4o, embed-and-save with OpenAI's embeddings). Here they are one idempotent pass the Worker runs on its own bindings (POST /api/index, lib/ai/index-images.ts), and every deploy's smoke step runs it: images in the bucket it has not seen are described and embedded, images it has seen are skipped by their R2 version and content hash, images gone from the bucket leave the index. The models are Workers AI's (lib/ai/models.ts) — Llama 4 Scout for the descriptions, BGE base at 768 dimensions for the embeddings — so the platform brokers no AI vendor and the Cloudflare account the app deploys into is the only one. Upstream ships no images; a fresh product starts with twelve drawn for this baseline, licensed with it, seeded into its bucket by its first pass. What the baseline adds is everything needed to make a copy of it somebody else's — built the way cirrus-baseline is. See specs/epics/semantic-image-search for the design, the milestones and the as-built status, including every departure.

Instantiating a product

git clone --depth 1 --branch <baseline-tag> https://github.com/sourceplane/semantic-image-search
cd semantic-image-search
orun new --blueprint repo-blueprint.yaml --out ~/products/acme-gallery --run-hooks --resume \
  --set reponame=acme-gallery --set productname="Acme Gallery" \
  --set githuborg=acme-inc --set subdomain=<workers-dev-subdomain>

Or, once the platform's registry lists it, from the Orun console (Baselines → semantic-image-search → Build). Two phases: 01-product creates the repository, mints the six brokered credentials (Cloudflare and Supabase) and lands the whole product as one pull request, whose convergence on main provisions and deploys it per environment in a fixed order —

supabase ──▶ cloudflare-hyperdrive ──▶ db-migrate ──▶ web
                        cloudflare-r2 ──────────────▶ web
                        cloudflare-kv ──────────────▶ web

— each edge mode: enforced on a push to main and advisory on a pull request — and the web lane's smoke indexes the images; then 02-docs records the live deployment from probed facts. BOOTSTRAP.md is the operator guide; docs/phases/ has a page per phase.

It creates two Supabase projects, <reponame>-stage and <reponame>-prod. Supabase's free plan allows two active projects per organization, so the connected organization needs room for both (BOOTSTRAP.md, §0).

The new repository receives the product only — the app, its sample images, the Terraform roots, the migrations, the tests, its CI, its tooling, a README of its own and its operating contract — renamed throughout to the product's name, repository, Worker, bucket, namespace, Supabase project and Hyperdrive config. None of this repository's machinery ships. No AI vendor key is asked for: Workers AI needs none, and bills the Cloudflare account the Worker runs in after its free daily allocation.

What is here (the product)

Getting started

pnpm install
pnpm run wire:fixture   # re-render apps/web/wrangler.jsonc from its template (committed; this changes nothing)
pnpm typecheck
pnpm lint
pnpm test
pnpm build              # next build + opennextjs-cloudflare build → apps/web/.open-next
pnpm --filter @saas/db run manifest:check

The build needs no database, no model, no bucket and no Cloudflare credential. Running it locally — a Postgres with pgvector on localhost:5432, pnpm --filter @app/web db:migrate:local, wrangler login for Workers AI, wrangler dev — is in the product's README (.baseline/product/README.md), which is what a product built from here carries.

How it maps to Vercel

CI