Ce que nous avons livré

Problème

Every AI-engineering brief in our inbox in late 2025 asked for some form of retrieval-augmented generation. The available open-source starting points were either toy demos or sprawling frameworks (LangChain) that bring more cognitive overhead than the actual RAG work. We needed a reference implementation that was production-shaped — auth, error handling, observability, real costs — but small enough that an engineer could read it end-to-end in twenty minutes.

Approche

We mapped the five RAG primitives — chunking, embedding, vector storage, source storage, generation — to one Cloudflare service each. Chunking is plain TypeScript in the Worker. Embedding is Workers AI (`@cf/baai/bge-base-en-v1.5`, 768-dim, free at low volume). Vector store is Vectorize. Source text is D1. Generation is Anthropic Claude Sonnet 4.6, streamed back to the client. No glue, no framework, no abstraction layers we did not write ourselves.

Résultat

Shipped as cf-rag-template under MIT. Deploys in under five minutes to any Cloudflare account. Has a companion long-form essay explaining every line and the cost model. Used internally as the starting point for every RAG-shaped client engagement since.

Problème

Model Context Protocol was the most-asked-about Anthropic primitive of 2025. The MCP starter repos floating around in late 2025 were demo-shaped — they showed the JSON-RPC handshake, but skipped bearer-token auth, structured logging, JSON-RPC error codes, request validation, a landing page for humans who hit the URL, and a health endpoint for the operator on call at 2 a.m. We needed a reference that included the production hygiene from line one.

Approche

Built a Cloudflare-Workers-native MCP server that speaks JSON-RPC 2.0 over HTTP POST (the modern MCP transport), authenticates via a bearer token stored as a Worker secret, and exposes one sample tool (`getCompanyFacts`) that downstream forks replace with their own. Every error path returns a proper JSON-RPC error code; every request is structured-logged; the root path serves a human landing page so direct visitors see something meaningful.

Résultat

Shipped as cf-mcp-template under MIT. Drop-in compatible with Claude Desktop, Cursor, and any MCP client. Companion essay covers the operating model (bearer rotation, request rate limits, what to log without leaking). Used internally as the starting point for every MCP-shaped engagement.