Introducing Ask AI, the reader assistant
Aardvark 0.1.6 shipped a native Ask AI panel that answers reader questions from your own docs, with cited sources and metered, dollar-based billing.
Since 0.1.6, every Aardvark site can ship a native Ask AI assistant. A floating panel sits on every page; readers ask a question in natural language, and the assistant answers from your content, citing the pages it used. Each answer can be rated 👍 / 👎, and — since 0.1.9 — readers can attach images, PDFs, or code when a screenshot says it better than a sentence.
There is no third-party widget to embed. The assistant is a first-party feature that calls the Aardvark cloud gateway, which proxies the model, meters the spend, and records the conversation for analysis. Enabling it is two lines of config:
ai:
assistant:
enabled: true
plus a spend-capped public gateway key (aardvark_live_…) baked in as the AARDVARK_KEY
build-time environment variable — never a provider key in your repo, and never the secret
key: a build handed one fails rather than shipping it inside a static site.
Two design choices matter most. First, grounding: when the corpus fits the browser’s 8 MiB safety
boundary, opening Ask AI loads the site’s agent-audience corpus, ranks each question locally, and sends
a small set of complete, structure-preserving Markdown pages in one model request. Larger corpora and
weak or ambiguous matches keep bounded page-by-page navigation; an oversized version-scoped index
fails closed rather than mixing releases. The whole corpus is never sent merely because a large model
window can hold it, and either path answers from what you actually published. Second, honest
metering: usage bills in real dollars against the account behind that key, so a public docs
site can offer AI answers without signing up for an unbounded bill. Point the panel at a
:free model slug and it answers at $0 while you evaluate it.
Behind the panel sits a conversation-analytics dashboard — top questions, coverage gaps, intent tags — so every reader question becomes product insight. You’re reading the docs of a site that runs it: press Cmd+I and ask something.
Read more on the AI assistant page, or see the changelog entry for the release details.