Bots and knowledge#
tufu has two building blocks:
- Knowledge: what you want tufu to know. A knowledge source is a website, a set of uploaded files, a GitHub repo, a Confluence space or your Slack history.
- Bots: what your visitors talk to. Each bot answers from the knowledge sources you give it, in a support chat or Ask console on your site, in Slack, and by email if you forward your support address to it.
A bot can use several knowledge sources, and a knowledge source can be used by several bots. For example, a shop with a public help center and an internal wiki could have:
- a public support bot that uses only the help center, so private information never reaches visitors;
- an internal bot for the team that uses both.
Nothing is copied twice: both bots read the same help center.
How an answer is written#
- A visitor asks a question, say "How long does shipping to Canada take?"
- tufu searches your knowledge for the passages most likely to answer it, like a search engine over your own docs. The search runs on tufu's own search model, whatever model your bot answers with.
- The AI model reads those passages along with the question and writes a short answer.
- The answer links to its sources, the pages it came from, so the visitor can read more and you can check its work.
Only the passages that matched are sent to the model, never your whole knowledge base. By default the model is tufu's AI, included in your plan. On either plan you can use your own AI key and pick the model yourself: see Bring your own model.
If your docs don't cover a question, the bot's Use general knowledge setting decides what happens. On (the default), the bot can also answer from what the model knows in general. Off, it only answers from your knowledge and says when it can't help. Either way, in a support chat the visitor can ask for a person: see Human handoff and the inbox.
What tufu is good at#
tufu shines when the answer is somewhere in your docs and the visitor doesn't know where:
- product questions: shipping, sizing, returns, custom orders, pricing;
- how-to questions about your software or service;
- finding the right page in a large help center.
What it's not good at#
Because the bot only ever reads the passages that matched, it struggles with questions about your docs as a whole:
- "Summarize everything you know";
- "What topics do you cover?";
- calculations across a spreadsheet.
A good rule of thumb: questions whose answer is in the content work well; questions about the content work less well.
Getting better answers#
- Write it down. Most wrong answers mean the answer isn't in your knowledge yet, or it's buried. A short, clearly titled page per common question works best.
- Keep knowledge fresh. Rebuild a source after you change your docs (see Adding knowledge).
- Use extra instructions for tone, not facts. "Keep answers short" or "Always reply in Spanish" belong in the bot's Extra instructions. Facts belong in your docs, where every bot can use them.
- Try it in Preview. Ask the questions your customers ask before they do.