DocFlare AI packages the plumbing behind “ask our documentation” into one Cloudflare Worker. Its initial source implementation, version 0.1.0, was committed on October 8; there is no tagged release yet. It is aimed at developers maintaining public documentation who would rather operate an inspectable integration than subscribe to a hosted chatbot service.

  • Retrieval before generation: A sitemap crawler extracts HTML and splits it at headings. Workers AI creates search embeddings; Vectorize retrieves 20 candidate passages, a reranker re-scores them, and Llama 3.1 8B generates an answer with source links. This is retrieval-augmented generation: the model receives relevant excerpts, not the entire website.
  • Setup is more than the embed: Use Bun and Wrangler, create a D1 database and a 384-dimensional cosine Vectorize index, apply the database migration, set an ADMIN_TOKEN, and deploy. Register your sitemap through /admin or POST /api/sites/index, then add the widget.js script with data-site-id to your docs. The deploy-button path still needs the database migration afterward.
  • Interface and operations: POST /api/chat returns JSON or a streaming Server-Sent Events response. Indexing runs in resumable batches; the defaults are three pages per batch and a 200-page sitemap cap. D1 stores questions and answers for the admin log, so this is not a no-retention service.
  • Cost: The code is MIT-licensed. Cloudflare’s Workers AI pricing includes 10,000 compute “neurons” daily at no charge; exceeding that requires the Paid plan and costs $0.011 per 1,000 additional neurons. Vectorize, Workers, and D1 have separate quotas. “Free” means staying within all of them—not unlimited chat. The repository’s traffic estimates are estimates, not measurements reproduced here.
  • Important boundary: There is no built-in authentication for private documentation. The chat endpoint accepts requests without an Origin header, so its browser-origin check is not an access-control boundary. Use public content only unless you add real authentication. The crawler also requires server-rendered HTML, and the default embedding model is English-oriented.

Worth considering for a small public docs site already on Cloudflare. Less compelling for a private knowledge base or an offline/self-hosted workflow: inference and vector search still depend on Cloudflare, even in local development. This is a source-and-documentation review, not a deployed trial or an independent answer-quality evaluation.