Documentation chat

Dory can add an Ask Dory assistant to your site. It is optional. The browser sends questions to a same-origin /api/chat endpoint; only the server calls your model provider. Answers include links to the documentation sections used as sources. When the docs have no matching source, the assistant says it cannot find the answer.

Configure a model#

Set these environment variables on the server running dory dev or dory preview:

DORY_CHAT_API_KEY=your-server-only-key
DORY_CHAT_MODEL=gpt-6-luna
# Optional: an OpenAI-compatible provider endpoint; defaults to https://api.openai.com/v1
DORY_CHAT_BASE_URL=https://api.openai.com/v1

If your existing .env uses OPENAI_API_KEY, Dory accepts that as a fallback. Set DORY_CHAT_MODEL as well; no model is chosen silently.

The chosen provider must support the OpenAI-compatible /chat/completions API and max_completion_tokens. Keep the key in your server's secret manager or an uncommitted .env file. Do not use a VITE_ prefix for the key: Vite exposes variables with that prefix to browser code. Dory never puts DORY_CHAT_API_KEY in the generated site or sends it to the browser.

Run dory dev while writing, or dory build && dory preview to serve the built site with chat. A static-file host cannot execute /api/chat; if you deploy dist/ to one, put a compatible server endpoint behind the same path. Dory hides the chat button when no backend is configured.

When enabled, Ask Dory appears in the documentation search dialog. Readers can switch to chat, choose a suggested question, or select the Ask Dory result above page matches to ask their current search query. Supported answers include citations and documentation-focused follow-up questions. If a question is outside the published docs, Ask Dory says so without attaching an unrelated source and suggests topics from your configured questions instead.

To customize the suggested questions for your site, add this optional section to dory.json:

"chat": {
  "suggestedQuestions": [
    "How do I install the product?",
    "What should I try first?",
    "Explain this page"
  ]
}

Without this section, Dory suggests questions based on your site name. Keep provider keys in the server environment, never in dory.json.

Limits and security#

The backend rejects cross-origin requests, oversized messages, long conversation histories, and requests beyond its limits. The defaults are deliberately conservative:

When using dory dev on localhost, 127.0.0.1, or ::1, Dory skips the per-minute and daily request quotas so you can iterate on chat locally. The concurrent-call guard still applies. dory preview and deployed servers keep the quotas below.

Each request has a 20-second provider timeout. A question can contain at most 600 characters, and at most six previous turns are accepted. The limits above apply per Node process; for multiple replicas, enforce shared limits at your reverse proxy or edge as well. Dory uses the connection IP, not an untrusted X-Forwarded-For header. If you put the server behind a reverse proxy, configure limits there using its verified client IP.

The server retrieves relevant MDX sections and gives only those excerpts to the model. It validates returned source IDs before showing an answer. Model output can still be imperfect, so keep examples and API facts reviewed and use dory check in CI.