p10node/docflare-ai is an open-source chatbot widget project for documentation sites: by attaching a single script tag to a docs site, it answers questions and shows the supporting documents alongside them. The 2026-10-08 commit 6c00950c added the Workers backend, crawler, vector search, RAG, and admin UI all at once. All 29 changed files are newly added, with 4501 lines added and 0 lines removed.
The answering method is retrieval-augmented generation (RAG, i.e. a structure in which the model produces an answer grounded in the documents it retrieved). The README explains that Hono is mounted on Workers, the question is embedded in Workers AI with bge-small-en-v1.5, the top 20 results are found in Vectorize by 384-dimensional cosine distance, the ranking is then reranked with a cross-encoder called bge-reranker-base, and only the final top passages are handed to llama-3.1-8b-instruct-fp8 to produce the answer. Search results are split by per-site namespace, and site, page, and query records are stored in D1 (SQLite). The widget is pure JavaScript using Shadow DOM and is under 10KB, and the admin screen at the /admin path is a single static HTML file with no build step. The license is MIT.
The README states that all components run within the Cloudflare free tier and also lists the limits in a table: Workers 100,000 requests/day, Workers AI 10,000 neurons/day, Vectorize 5 million stored dimensions and 30 million dimensions queried per month, D1 5GB and 5 million reads per day. The crawler handles sitemap indexes and gzipped sitemaps, and strips out common elements such as navigation, headers, and footers with HTMLRewriter.
That said, these figures are the values written in the repository's documentation, not externally confirmed measurements. The README does not clarify how well an 8B model can be handled within these limits. This commit is not a feature fix but the addition of the entire source, and there is no separate version release notation. A unit test file, test/unit.test.ts, came in along with it, but the execution results are not preserved in this commit's information. Before deployment, the ADMIN_TOKEN placeholder in .dev.vars.example must be replaced with an actual value. No public benchmarks for search accuracy or response quality are provided, so the verifiable range comes close to no more than that this is an initial implementation worth trying on a documentation site.