An AI assistant is only as good as what it knows. The knowledge base is where that knowledge lives, and it is the single biggest factor in whether your assistant gives accurate answers or makes things up. Here is what to put in it, how it works, and how to get it right.
What a knowledge base is
A knowledge base is the set of documents and content your assistant can draw on: your website, service descriptions, pricing, policies, FAQs, hours, and locations. The assistant references this content to answer questions instead of guessing.
Why it matters for accuracy
Give the assistant little, and it improvises. Give it clear, well-organized content, and answers stay grounded in your real information. A strong knowledge base means:
- Better accuracy: answers come from your content, not the model's assumptions
- Less back and forth: customers get a complete answer the first time
- A consistent experience: the assistant sounds like your business, every time
How retrieval works
When someone asks a question, the assistant runs a semantic search across the knowledge base. Semantic search matches meaning, not exact keywords, so 'are you open Sunday?' and 'weekend hours?' find the same answer. The assistant then summarizes the relevant content and can cite the source.
What to put in it
Start with your highest-impact content: pricing, policies, services, onboarding, and your existing FAQ pages. Anything a customer commonly asks should be in here, named clearly so the assistant knows where to look.
Supported file types
- PDFs: manuals, guides, policies
- Excel or CSV: product lists, data tables
- Images: product photos, menus, catalogs (text is extracted)
- Audio files (up to 20 MB)
- Presentations and slides
- Plain text: paste a pricing list or any content directly
- Website links: index pages directly
Invent's knowledge management view: add documents, files, and links the assistant references when answering.
Knowledge base or actions?
A knowledge base is ideal for informational questions. For anything that must be exact, like multi-variable pricing or booking rules, an action that looks up structured data is more reliable. Most assistants use both. See how to train an AI assistant on your own data for the full breakdown.
How to test it
Ask your ten most common questions in different wordings and check the answers are consistent and correct. Ask about something you do not offer to see whether it stays grounded or invents a service. If answers drift, the fix is usually clearer document names or a missing action.
No real data yet?
- Generate sample user profiles to simulate conversations
- Ask a model to list the most common questions in your field
- Draft FAQ content from your existing site copy
FAQs
How much content do I need to start?
Enough to cover your top questions: pricing, policies, services, and FAQs. You can expand it as you see what customers actually ask.
What happens if the assistant does not know?
With a clear knowledge base it will say so and hand off to a person rather than guess. Configure handoff so nothing falls through.
Does a bigger knowledge base always help?
Clarity beats volume. Well-named, current documents outperform a large pile of overlapping or outdated files.
Common mistakes to avoid
- Dumping everything in with vague file names, so the assistant cannot tell which source is authoritative
- Leaving outdated docs in place, so the assistant confidently quotes old prices or policies
- Keeping pricing only inside marketing prose instead of a clear, named document or an action
- No handoff path, so when the assistant does not know, the conversation dead-ends
How often should I update the knowledge base?
Whenever pricing, policies, or services change. Stale content is the most common cause of wrong answers.
Can it use my website directly?
Yes, you can index website pages. Audit and remove outdated pages first so the assistant does not learn old information.
Ready to build one? Create your assistant with Invent, and see the step-by-step build guide.
