A full-stack RAG assistant built and deployed on Cloudflare’s own platform, as a tailored application artifact for Cloudflare roles. Live demo above (if reachable); fallback link below.

Stack

• Routing / orchestration: A Cloudflare Worker (TypeScript) that handles the inbound request, embeds the query (@cf/baai/bge-small-en-v1.5), queries the Vectorize index, assembles the prompt, and forwards to the Workers AI generation model.
• Knowledge memory: Cloudflare Vectorize — embeddings of Cloudflare’s own docs serve as the retrievable knowledge base.
• Conversational memory: Cloudflare Durable Objects — each user session is a distinct DO instance, so the assistant can carry multi-turn context without a global database.
• Generation: @cf/meta/llama-3-8b-instruct via Workers AI.
• Frontend: React + TypeScript on Cloudflare Pages (main.cloudflare-ai-assistant.pages.dev).

The memory architecture

The non-obvious design choice was to split memory along two axes that scale and fail independently:

• Knowledge memory lives in a Vectorize index. It’s append-only, read-mostly, high-cardinality. Embeddings are computed once and re-used across all users.
• Conversational memory lives per-user in a Durable Object. It’s mutable, read-write balanced, low-cardinality (1 DO per active session).

Conflating these would force every user’s transient chat state to scale with the size of the knowledge corpus — wrong shape for the workload, wrong cost curve.

Why this is on the portfolio

When I apply to Cloudflare, the implicit ask is “have you actually used what we build?” This project answers that with a deployed URL.