AI chatbot feature overview

Use our AI Chatbot to answer your readers' questions with information from your knowledge base. No separate accounts, complex integrations, or data import required.

The quick summary

Our AI Chatbot runs entirely within KnowledgeOwl’s infrastructure. Your data is never shared externally and is not used to train AI models.

Your readers ask a question of the chatbot. It analyzes information in your knowledge base--and ONLY in your knowledge base--and returns an answer using the tone and audience info you provide in your setup. It will include images from your documentation if they're relevant to answering the question.

The chatbot cites the sources it used, encouraging your readers to learn more in your knowledge base and establishing trust and authority.

And it admits when it doesn't know.

The more detailed process

Before enabling the chatbot on your knowledge base, we need to make sure your knowledge base's search can handle natural language queries (full sentence questions rather than keyword searches). To handle those natural language queries, we need to turn on semantic search indexing in your knowledge base search settings and then reindex your content to generate the semantic search index.

After the reindex, your content still has the same great keyword search capabilities it always has, but it also gains the critical natural language capability the chatbot needs. This gives you the best of both worlds. And importantly, both keyword and semantic serach indices update every time you make an edit, ensuring accuracy in its responses.

When a reader asks a question in the chatbot, we lean on the semantic side of your search index and send the top results from that search into the AI model to try to find an answer.

If the AI model has a high enough confidence that it can answer the reader's question using the article content that we've given it, it generates an answer using the prompt descriptions you entered in the chatbot settings for your company/audience and tone. 

The chatbot also includes hyperlinks to the articles used to generate the answer, so your readers know where to go to learn more and can assess whether it feels like the chatbot properly understood the question.

We try to ensure that if the AI model doesn't have a high confidence in answering the question using the article content we gave it, it says it can't answer the question (instead of trying to pull the answer out of thin air or its own existing memory, which is where most hallucinations come from).

Accuracy and security safeguards

You've probably heard a lot about Generative AI Large Language Models (LLM) offering incorrect information--known as hallucinations. So have we. We're firm believers that the information you store in KnowledgeOwl is yours. We know how much trust you place in us to be good custodians of that information, and we take that responsibility seriously.

We focused on designing a chatbot we felt confident using ourselves, one that could answer questions effectively and accurately while keeping our information secure.

Here's what that means in practice:

  • The chatbot doesn't try to "learn". We don't send information back to any model providers to train their models and we don't try to further train the model ourselves. The model we use is completely static.
  • Once the AI model generates its response, it doesn't store any history of the interaction. The question that was asked, the search results we provided, and the response it generated cannot be accessed again or used as any kind of reference for future questions.
  • Even when a reader asks multiple questions within the same session, they are treated as discrete, standalone questions. This allows us to keep your data secure not only from other knowledge bases and accounts, but also between readers who belong to different groups who can view different content.
  • Your knowledge base information is as secure in this process as it has always been.
  • You can learn from your AI chatbot. While we don't store any of the questions or articles in the AI model, we do store a history of the questions your reader asked the chatbot in our AI chatbot reporting. This is stored in the same way as our other Reporting data. This reporting can help you find cases where the chatbot wasn't confident enough to answer or to discover trends in the kind of information your readers are asking the chatbot for.