Knowledge base

Every article in your workspace knowledge base is split into paragraph-sized chunks. Each chunk gets a vector representation (an "embedding") stored alongside the article. When a visitor asks something, we embed the query, find the most similar chunks, and inject them into the AI agent's prompt next to the conversation history.

The AI is instructed to cite the article it leaned on for each part of its answer. If nothing relevant is found, it says so explicitly rather than fabricating an answer.

Three ways to populate the knowledge base:

• Write inline. Click Knowledge → New article and use the markdown editor. Best for short policies, FAQs, and response snippets the AI should reuse verbatim.
• Paste markdown. Bring existing help-site content as .md files. Multi-file import is supported under Knowledge → Import.

On each AI turn we retrieve the top-K most similar chunks (default K = 6). Chunks below a similarity floor are dropped. The remaining chunks are injected into the prompt with their source URLs.

Replies cite the source articles by title. Visitors don't see vector scores — they see a link to the canonical doc when the AI has paraphrased.

Edits to an article re-embed only that article. A full reindex (under Knowledge → Settings → Reindex) re-embeds everything — use this after a bulk content overhaul or if the AI starts citing stale articles.

See AI agent for how the AI actually uses what you put here.

Embeddings are produced by a platform-managed model — we handle the provider relationship so you don't have to shop for an API key. Vectors are stored per-workspace and isolated from every other workspace; nothing in your knowledge base reaches another customer's AI.

Knowledge-base content is treated the same way as conversation content for privacy — see Privacy Policy.