cf-rag-template

Production RAG pipeline on Cloudflare Workers. ~200 lines of TypeScript. No LangChain, no framework. Deploy in 5 minutes.

Companion to: Production RAG on Cloudflare Without LangChain

What it is

The five primitives of retrieval-augmented generation, mapped to one Cloudflare service each:

Step	Service	Cost at low volume
Chunk	Workers (your code)	Free
Embed	Workers AI (@cf/baai/bge-base-en-v1.5, 768-dim)	Free / trivial
Store vectors	Vectorize	$0.04 / 1M vectors / month
Store source text	D1	Free at this scale
Generate	Anthropic Claude Sonnet 4.6 (streamed back to client)	$3 / $15 per 1M tokens

One Worker, three bindings, one external API call. That's it.

What you get

Three endpoints:

• POST /ingest — { docId, text } → chunks, embeds, stores. Returns { chunks: N }.
• POST /ask — { query } → SSE stream of Claude's grounded answer with [1], [2] citations.
• POST /forget — { docId } → removes the document and all its vectors.
• GET /health — liveness check.

Setup (one-time, ~3 minutes)

# 1. Install
npm install

# 2. Create a D1 database
npx wrangler d1 create rag-db
# Paste the returned database_id into wrangler.jsonc

# 3. Run the migration
npm run db:init:prod

# 4. Create the Vectorize index
npx wrangler vectorize create rag-index --dimensions=768 --metric=cosine

# 5. Set the Anthropic API key as a secret
npx wrangler secret put ANTHROPIC_API_KEY
# (paste your key from https://console.anthropic.com/settings/keys)

# 6. Deploy
npm run deploy

The Worker is now live at https://cf-rag-template.<your-subdomain>.workers.dev.

Try it

WORKER=https://cf-rag-template.<your-subdomain>.workers.dev

# Ingest a document
curl -X POST "$WORKER/ingest" \
  -H 'content-type: application/json' \
  -d '{"docId":"doc-001","text":"The capital of Nova Scotia is Halifax. Founded in 1749, it sits on the second-largest natural harbour in the world."}'
# → {"chunks": 1}

# Ask a question — streams the answer
curl -N -X POST "$WORKER/ask" \
  -H 'content-type: application/json' \
  -d '{"query":"Where is the capital of Nova Scotia?"}'
# → SSE stream with Claude's response, citing [1]

# Forget the document
curl -X POST "$WORKER/forget" \
  -H 'content-type: application/json' \
  -d '{"docId":"doc-001"}'
# → {"removed": 1}

How to extend

• Better chunking — replace the naive sliding window in chunk() with a corpus-aware splitter (semantic chunking, heading-aware). Single biggest lever on retrieval quality.
• Different embeddings — swap @cf/baai/bge-base-en-v1.5 for @cf/baai/bge-large-en-v1.5 (slightly better quality) or OpenAI's text-embedding-3-small / -large. Remember to update --dimensions on the Vectorize index.
• Different model — claude-sonnet-4-6 is a sweet spot. For cheaper/faster: claude-haiku-4-5. For long context (1M tokens): claude-opus-4-7. For routing per task, see the Workers AI vs OpenAI matrix.
• Auth — the worker is open by default. Add a bearer-token check on the entrypoint before deploying to production.
• Hybrid retrieval — pair Vectorize with a lexical-ranker (BM25 in D1) and merge results for higher recall.
• Eval harness — write a list of 30–100 known queries with expected answers; run them on every change. Track precision/recall. This is the single most important piece a framework can't give you.

What this isn't

• Not a framework. Don't use it as a dependency. Fork it, read it, change what you need.
• Not opinionated about the chunking strategy — that's your corpus's decision.
• Not auth'd. Add auth before exposing to the internet.

License

MIT — see LICENSE.

Built by

SetKernel Digital Inc. — a Cloudflare-native engineering studio. We design, build, and operate AI-augmented products on the edge. Need a production RAG system that actually ships? Write a brief.