AI configuration and providers

Hark's AI features (the chat bot, AI flow nodes, semantic search, and the agent copilot) run through an AI provider you connect. You bring your own provider, so you control the model and the cost.

Providers

You configure one AI provider per workspace. Supported providers include OpenAI, Anthropic, an OpenAI-compatible endpoint, a local LM Studio server, and Ollama. For each you set the base URL where needed and an API key, which is encrypted before it is stored.

Models

You choose two models:

• A chat model for generating and classifying text (bot answers, copilot responses, flow AI nodes).
• An embedding model for semantic search. Embeddings are fitted to a fixed dimension so search stays consistent across content.

The handoff threshold

The AI carries a confidence handoff threshold. When the AI's confidence in an answer falls below it, the conversation hands off to a human rather than guessing. Raising the threshold makes the bot more cautious (more handoffs, fewer risky answers); lowering it lets the bot attempt more on its own.

Semantic search

Your public corpus (help articles, feedback posts, and changelog entries) is indexed into a vector store so AI features can retrieve relevant context by meaning, not just keywords. The index is maintained in the background and only re-embeds content that changed, so it stays fresh without redundant work.

Honest degrade

Everywhere AI is used, the design is the same: if the model is unavailable or unsure, Hark degrades honestly (a flow takes its failure branch, the bot hands off, search falls back to keywords) rather than fabricating a result. AI is a visible, accountable layer, not a black box.