Personal AI Chatbot — RAG on Cloudflare Edge

This project is an LLM-powered chatbot with a custom RAG system, deployed on Cloudflare's edge network. Check out the live demo at chienliu.com — the chatbot can answer questions about the author, Chien Liu 😺

• Cloudflare Workers: Edge runtime hosting the chatbot API
• Workers AI: LLM inference for chat responses
• D1: Relational store for RAG note metadata
• Vectorize: Vector store for RAG embeddings/retrieval
• Workflows: Orchestrates the RAG note sync process
• Hono: Lightweight router for the Worker

The chatbot supports multi-turn conversations. It can answer follow-up questions like "Elaborate on that" or "What else?" because each request optionally carries an array of prior messages.

For a simple personal chatbot, frontend storage is the better call. Leveraging frontend storage is simpler, cheaper, and more privacy-compliant.

• This repository is integrated with Cloudflare Workers.
• The deployed Worker is built from the code in src/, with src/worker.js as the entrypoint.
• The production route is https://api.chienliu.com/chatbot.
• The local maintenance assets in notes/ and scripts/ support RAG content management, but they are not the Worker code that gets deployed from src/.

• POST /chatbot is rate limited per client IP using the native Cloudflare Workers Rate Limiting binding (CHATBOT_RATE_LIMITER, configured in wrangler.jsonc).
• Limit: 5 requests per 60 seconds per IP (identified via the CF-Connecting-IP header).
• Requests over the limit receive 429 Too Many Requests.

• Each markdown file in notes/ is the source of truth for one RAG note.
• The filename must follow this format:

${id}_${name_for_human_editor}.md

Examples:

1_chiens_work_experience.md
2_projects.md
15_publications.md

• ${id} must be a positive integer.
• The sync script normalizes the filename ID before sending it to the API.
• The resulting ID is reused as the note ID in D1 and Vectorize.
• Remote note writes happen through PUT /notes/:id (local only) with a JSON body shaped like { "text": "..." }. In production, only /chatbot requests reach the Worker, so /notes/:id is unreachable at the edge.
• Renaming or deleting local files affects remote data on the next sync.

1. Create or edit markdown files in notes/.
2. Copy .env.local.example to .env.local (first time only):

cp .env.local.example .env.local

3. Start the local Worker:

npx wrangler dev

4. In another terminal, run the sync:

npm run sync:notes

The workflow may take a moment to process. Monitor the npx wrangler dev output to confirm the upload has completed.

The sync script is scripts/sync-rag-notes.sh.

It will:

1. Scan notes/ for *.md files.
2. Validate that every filename matches ${id}_*.md.
3. Compute a SHA-256 hash for each file.
4. Upload only notes whose content changed since the last successful sync.
5. Delete remote notes whose IDs existed in the previous sync state but no longer exist locally.
6. Save sync state in .rag-sync-state.

Because of this behavior:

• Editing a file updates the remote note
• Adding a file creates a remote note
• Deleting a file deletes the remote note
• Changing only the human-readable suffix keeps the same remote note if the numeric ID stays the same

• Missing env variable: ensure .env.local exists and contains ADMIN_API_ENABLED=true.
• Notes directory not found: create notes/ or set RAG_SYNC_NOTES_DIR.
• Invalid note filename: rename the file to ${id}_name.md.
• Duplicate note id: ensure only one markdown file uses each numeric ID.

wrangler dev in this project is configured with remote Cloudflare bindings. A live sync updates the bound D1 and Vectorize resources, so use the sync flow carefully.