AI Chat

OwnDocs ships an AI chat widget that answers questions about your documentation and cites the pages it uses. The widget stays hidden until you configure a provider, so there is nothing to turn off when you do not want it.

How it works

The widget is a fixed button in the bottom-right corner. It sends the reader's question to /api/ai/chat, which retrieves the most relevant pages from the search index, builds a context window, and streams an answer with source citations. Chat threads are stored locally in the reader's browser.

Prerequisites

• A provider: Anthropic, an OpenAI-compatible endpoint (OpenAI, OpenRouter, Moonshot Kimi, Z.ai GLM, DeepSeek, Groq, Mistral, xAI, and more), or Ollama (local or cloud).
• The environment variables for that provider, set before building.

Enable with Anthropic

Anthropic is the default when no AI_PROVIDER is set.

ANTHROPIC_API_KEY="your-anthropic-key"
ANTHROPIC_MODEL="claude-sonnet-5"

Enable with an OpenAI-compatible endpoint

AI_PROVIDER="openai-compatible"
AI_API_KEY="your-api-key"
AI_BASE_URL="https://api.openai.com/v1"
AI_MODEL="gpt-5.6-luna"

For OpenRouter, point AI_BASE_URL at https://openrouter.ai/api/v1 and use an OpenRouter model id.

More OpenAI-compatible providers

Any endpoint that speaks OpenAI's chat completions format works here. A few more, each with a working example.

Moonshot AI (Kimi)

AI_PROVIDER="openai-compatible"
AI_API_KEY="your-moonshot-key"
AI_BASE_URL="https://api.moonshot.ai/v1"
AI_MODEL="kimi-k3"

Do not set temperature — Kimi K3 fixes it at 1.0 and rejects any other value. Accounts on Moonshot's mainland-China platform use https://api.moonshot.cn/v1 with a separate key instead.

Z.ai (GLM)

AI_PROVIDER="openai-compatible"
AI_API_KEY="your-zai-key"
AI_BASE_URL="https://api.z.ai/api/paas/v4"
AI_MODEL="glm-5.2"

Z.ai is the current name for what was Zhipu AI. Use glm-5.2, not the newer glm-5.3 — as of this writing GLM-5.3 is available only through Z.ai's subscription coding plan, not the pay-per-token API.

DeepSeek

AI_PROVIDER="openai-compatible"
AI_API_KEY="your-deepseek-key"
AI_BASE_URL="https://api.deepseek.com"
AI_MODEL="deepseek-v4-pro"

The older deepseek-chat and deepseek-reasoner model ids were retired. Use deepseek-v4-flash for a faster, cheaper tier of the same model family.

Groq

AI_PROVIDER="openai-compatible"
AI_API_KEY="your-groq-key"
AI_BASE_URL="https://api.groq.com/openai/v1"
AI_MODEL="openai/gpt-oss-120b"

Groq does not support the logprobs, logit_bias, or top_logprobs parameters, and treats temperature: 0 as a very small positive number instead of exactly zero.

Mistral AI

AI_PROVIDER="openai-compatible"
AI_API_KEY="your-mistral-key"
AI_BASE_URL="https://api.mistral.ai/v1"
AI_MODEL="mistral-medium-latest"

mistral-medium-latest always points to Mistral's current mid-tier flagship model. Pin an exact version with mistral-medium-3-5 if the response needs to stay consistent over time.

xAI (Grok)

AI_PROVIDER="openai-compatible"
AI_API_KEY="your-xai-key"
AI_BASE_URL="https://api.x.ai/v1"
AI_MODEL="grok-4.6"

xAI is now part of SpaceX and its documentation is branded SpaceXAI, but the API endpoint, model ids, and XAI_API_KEY variable name are unchanged.

Enable with Ollama (local)

Local Ollama needs no API key.

AI_PROVIDER="ollama"
AI_MODEL="gemma4:12b"

For a smaller download on memory-constrained machines, use gemma4:e2b or gemma4:e2b-it-qat instead.

Enable with Ollama Cloud

AI_PROVIDER="ollama-cloud"
AI_API_KEY="your-ollama-cloud-key"
AI_MODEL="deepseek-v4-flash:0731"

Ollama Cloud's model ids do not follow one fixed pattern: through the local ollama CLI, cloud models take a :cloud tag (deepseek-v4-flash:cloud); called directly at the https://ollama.com/v1 endpoint shown above, they use the plain dated id instead, with no suffix. Check Ollama's cloud model list for the current id before deploying, since cloud models are retired and replaced on their own schedule.

Provider options

Result

After setting a provider and rebuilding, the chat button appears in the bottom-right corner of every documentation page. Ask a question about the docs and the answer lists the source pages it used.

Troubleshooting