AI Knowledge Base and Training

The knowledge base is the set of source material an AI agent is allowed to draw on when it answers. You add documents, pasted text, question-and-answer pairs and website pages; the platform indexes them; and when a contact asks something, the agent retrieves the relevant passages and answers from them rather than from general knowledge.

This is the difference between an agent that sounds plausible and an agent that is correct. Instructions in AI Studio control how the agent behaves. Knowledge controls what it knows. An agent with excellent instructions and no knowledge will be polite and wrong about your client's hours, services and policies, because it has nothing to consult and will not volunteer that it does not know unless you told it to.

Treat it as a reference shelf, not an archive. The useful mental model is a binder behind the front desk containing exactly the answers to the questions people actually ask, written the way you would want them said aloud. Operators get the worst results by dumping an entire website and a folder of PDFs in and hoping relevance sorts itself out. The best results come from a small, deliberately written set of answers that covers the real question distribution, kept current, with everything else deliberately left out.

Source types

• Uploaded documents such as service menus, policy sheets and price-free information guides
• Pasted text blocks written specifically for the agent
• Explicit question-and-answer pairs, which give you the tightest control over wording
• Crawled website pages, with control over which URLs are included
• Links to specific pages you want the agent to be able to send the contact to

Scoping

• Which sources are attached to which agent, so a sales agent and a support agent read different material
• Enabling and disabling a source without deleting it, which is how you handle seasonal or paused offerings
• Grouping or labeling sources so a large base stays maintainable

Retrieval behavior

• How strictly the agent must stay within the sources, versus being allowed to answer generally
• What it says when nothing relevant is found
• Whether it cites or links the source it used

Freshness and maintenance

• Re-crawl or re-index after a website or document changes
• Visibility of when each source was last indexed
• Review of unanswered or poorly answered questions so you can see the gaps

Testing

• Asking the agent questions directly against the knowledge base and seeing which passage it used
• Checking for contradictions between sources on the same topic

1. Start from the questions, not the documents. Ask the client for the questions their front desk answers every day, and get the real list, including the awkward ones.
2. Write the answers yourself, short, in the voice the client wants, and have the client approve them. These become your question-and-answer pairs and they will carry most of the load.
3. Add the operational facts the agent will otherwise invent: hours, locations, service area, what the client does and explicitly does not offer, how to reach a human.
4. Add the documents that genuinely contain answers. Convert anything that is a scanned image or a slide deck into clean text first, because retrieval from a messy file is poor.
5. If you crawl the website, scope the crawl. Include service and FAQ pages; exclude blog archives, legal boilerplate and anything with outdated offers.
6. Remove or exclude every source containing information the agent must never state, rather than instructing it not to mention something it can read.
7. Attach the sources to the specific agent that needs them. Do not share one undifferentiated base across agents with different jobs.
8. Set the strictness and the not-found behavior. For most client work the agent should stay inside its sources and say plainly when it does not know, then offer a human.
9. Index the sources and confirm each one shows as successfully processed. A silently failed upload is a common cause of an agent that suddenly knows nothing.
10. Test by asking the client's real questions, one at a time, and check which source each answer came from.
11. Hunt for contradictions: the old hours on a crawled page versus the new hours in your pasted block. Fix the source, do not patch it with an instruction.
12. Ask the questions you expect it to refuse and confirm it refuses rather than improvises.
13. Agree a maintenance owner and cadence with the client, and a rule that whoever changes the website or a policy also updates the knowledge base.
14. After go-live, review unanswered questions weekly for the first month and add the gaps as new question-and-answer pairs.

Knowledge is consumed by every agent you build: the Conversation AI agent on text channels, the Voice AI agent on the phone, and AI review replies when they need facts about the business. It is authored and attached from AI Studio.

Clinic or service business front desk

The base holds locations, hours, what each service involves, what to bring, insurance or payment handling in general terms, and the policies on lateness and cancellation. The text and voice agents both read it, so a contact gets the same answer whether they call or text. Updating hours in one place updates both.

Seasonal or promotional offerings

A limited offering gets its own source. It is enabled when the promotion starts and disabled when it ends, so the agent stops offering it the moment it is over without anyone editing instructions. This is far safer than relying on the agent to notice a date.

Separate sales and support agents

One agent handles new inquiries with sources about services and booking. A second handles existing clients with sources about policies, rescheduling and support escalation. Keeping the bases separate stops the sales agent from reciting support policy and keeps each agent's answers tighter.

• More sources make answers worse, not better. A large base produces weaker retrieval because irrelevant passages compete with the right one. Curate aggressively.
• Contradictions are the most damaging failure. Two sources stating different hours produce an agent that is confidently wrong half the time, and you cannot tell which it will pick. Resolve at the source.
• Crawled sites go stale and so does the index. Updating the website does not update the agent until the source is re-indexed. Make that an explicit step in any site change.
• Scanned PDFs and image-heavy decks index poorly. If the text is not extractable, the agent cannot read it, and nothing will tell you except bad answers.
• Knowledge cannot be unsaid by instruction. If a source contains something the agent must not state, remove the source. Telling it not to mention what it can read is unreliable.
• An empty or failed index looks like a broken agent. Always confirm processing succeeded after upload, and after any bulk change.
• Anything priced or contractual ages fast. Prefer sources that describe what something is and route specifics to a human, rather than sources that quote terms that will change.
• Retrieval is not reasoning. The agent will answer the question it matched, not necessarily the question asked. Compound or conditional questions still need a handover path.
• Client-supplied material needs review. Internal documents often contain staff notes, old policies and commentary the client would not want quoted to a customer.
• Ownership gaps kill it slowly. A knowledge base with no named maintainer is accurate on launch day and misleading within a quarter.