DocDesk

Case study · Forward-deployed engineering

Dense docs.Instant, sourcedanswers.

Give DocDesk a company's domain. It finds their public docs — llms.txt, docs sites, sitemaps, OpenAPI specs, GitHub — and answers questions from them, citing every source. When the docs don't say, neither does DocDesk.

Ask it something See the architecture

01 — Live demo

Point it at any company

Enter a domain. DocDesk finds the company's public docs, shows you what's there, and builds a sourced assistant from the sources you pick.

  1. Domain
  2. What we found
  3. Chat
Try an example:

    02 — How it works

    Discover, index, ground, guard

    1. 1

      Discover

      Checks llms.txt, docs hosts, sitemaps, OpenAPI specs and GitHub — honouring robots.txt — and flags docs behind a login.

      /api/discover
    2. 2

      Index

      Pages are ranked by relevance to common integration topics and your focus, then the best 40 become ≤300 passages — embedded with bge on Workers AI and kept 6 hours in KV.

      /api/index
    3. 3

      Ground

      Workers AI writes a short answer using only those passages, citing each one. No match above the threshold means “I don't know.”

      llama-3.1-8b-instruct-fast
    4. 4

      Guard

      Per-session page caps, per-visitor rate limits and daily caps keep the demo inside Cloudflare's free allowance.

      < 10,000 neurons / day, by design
    DocDesk request flowVisitor question goes to a Cloudflare Worker, which checks the rate limit, retrieves passages from the index, sends them to Workers AI, and returns a cited answer. Visitordomain → question Workerdiscover · indexcaps · /api/chat Session indexKV · 6 h · top-5 Workers AIgrounded answer Answer+ citations
    Everything runs on one Cloudflare Worker — static site and API together.

    03 — Where it looks

    Public sources, in order

    1. llms.txt / llms-full.txt on the site and docs subdomain
    2. Docs hosts — docs., developer(s)., api., /docs, /developers (ReadMe, GitBook, Mintlify, Docusaurus detected)
    3. sitemap.xml, filtered to docs-like URLs
    4. OpenAPI / Swagger specs — endpoints grouped by tag
    5. GitHub org — OpenAPI spec repos, plus docs, API and SDK READMEs

    If the official docs are behind a login or a "contact your rep" wall, DocDesk says so and falls back to the rest. It never makes up a source.

    04 — Why I built it

    Customers shouldn't have to read the manual to get an answer.

    As a forward-deployed engineer, the first job on every new account is finding where the real docs live — llms.txt, a spec repo on GitHub, a README — and then answering the same questions over and over. DocDesk automates both: it maps a company's public docs in seconds and turns them into instant answers that link straight to the source.

    Built by Justin Son