Training an AI agent on your own documents: knowledge base setup
Training an AI agent on your own documents means uploading files or URLs to its knowledge base, which the agent searches through and quotes from when answering (a technique called RAG). Upload your FAQs, pricing sheet, policy documents and product details, then test with real customer questions and refine the content that gets answered poorly. Review the knowledge base every time your pricing or policies change.
- A knowledge base lets an AI agent answer from your actual documents instead of only its system prompt.
- Upload focused, well-organised documents — a clean FAQ works better than one giant unstructured PDF.
- The agent breaks documents into smaller chunks internally to search efficiently; you don't need to do this manually.
- Test with the exact questions real customers ask, not questions you assume they'll ask.
- Outdated documents cause outdated or wrong answers — review the knowledge base whenever pricing or policy changes.
Prerequisites and what a knowledge base actually does
You need an existing AI agent (see our system prompt guide if you haven't set one up) and your source documents gathered in one place: pricing sheets, FAQs, policy pages. Gathering these first, rather than uploading piecemeal as you think of them, makes it far easier to spot gaps or contradictions before they reach a customer. It also helps to have a short list of real customer questions ready, either from past chat logs or from your team's memory of what gets asked most often at the counter or on calls.
By default, an AI agent answers based on its system prompt: instructions you write about tone, rules and basic facts. A knowledge base goes further: you upload documents or URLs, and the agent searches through them for relevant content before answering, quoting from your actual material instead of guessing. This approach is often called RAG (retrieval-augmented generation); see our plain-language explainer if you want the underlying concept.
The practical difference customers notice is specificity. A system-prompt-only bot can say 'we offer a warranty on our products' in general terms; a bot with a knowledge base can say exactly what the warranty covers, for how long, and what voids it, because it's reading that detail from your actual policy document rather than paraphrasing a general instruction.
What to upload, and what to leave out
- Good candidates: a clean FAQ document, your pricing sheet, policy pages (returns, cancellation, warranty), product spec sheets, and service area details.
- Less useful: marketing brochures full of design flourishes and little factual text, very long documents mixing many unrelated topics, or outdated PDFs still floating around your drive.
A well-organised, focused set of documents outperforms one giant catch-all file. If you have a 40-page product manual, consider splitting out just the customer-facing FAQ section rather than uploading the whole manual, since the rest of the manual is unlikely to answer the kind of questions a WhatsApp customer actually types in.
How chunking works, in plain terms
You don't need to manually prepare your documents in any special way. The platform automatically breaks each uploaded document into smaller chunks (a paragraph or section at a time) so it can search and retrieve just the relevant part when answering a question, rather than reading the whole document every time. The one thing worth knowing: documents with clear headings and short, self-contained paragraphs chunk more cleanly than a dense wall of text, which tends to improve answer accuracy.
Exact click path: uploading and testing on Convo360
On Convo360, uploading and testing a knowledge base happens entirely inside the agent's own settings, with no separate tool to configure:
- Open the AI agent you want to train and go to its Knowledge Base section.
- Upload documents (PDF, Word, text) or paste URLs to relevant web pages.
- Wait for processing to complete; the platform indexes the content automatically.
- Open Preview / Try-it and ask the exact questions your customers actually ask, not hypothetical ones.
- If an answer is wrong or vague, check whether the source document actually covers that topic clearly, and edit the document if needed.
- Re-test after any document update.
A worked example: a jewellery store's knowledge base
A jewellery store in Coimbatore uploads three documents: a plain-language FAQ (making charges, hallmarking, exchange policy), a warranty policy page, and a one-page current-collection catalogue with indicative pricing ranges. During testing, the agent answers 'do you charge making charges on gold coins' correctly straight from the FAQ, but stumbles on 'can I exchange an old necklace for a new design' because the exchange policy document only mentioned exchanging within the same category. The owner adds one clarifying line to the policy document, re-uploads it, and the next test question gets answered correctly, a five-minute fix that would otherwise have meant an inaccurate answer reaching a real customer.
Troubleshooting common issues
| Issue | Fix |
|---|---|
| Agent gives a vague or generic answer | Check the source document actually covers that specific question clearly, not just the general topic. |
| Agent gives an outdated answer | Re-upload the corrected document, since old versions can linger if not explicitly replaced. |
| Agent contradicts itself across two questions | Look for conflicting information across two different uploaded documents and reconcile them. |
| Agent answers confidently but incorrectly (hallucination) | Add an explicit system prompt instruction to only answer from the knowledge base and say when unsure. |
| Processing seems stuck after upload | Very large or oddly formatted files can take longer to index. Wait a few minutes and refresh before re-uploading. |
Guardrails against hallucination, and keeping it current
Even with a knowledge base, an AI agent can occasionally state something confidently that isn't quite right. This is called hallucination. Reduce this risk by writing an explicit instruction in the system prompt: 'Only answer from the knowledge base; if you're unsure, say so and offer to connect the customer with a human.' Keep documents accurate and unambiguous, and set a lower temperature for agents that need to be precise and literal, rather than creative.
A knowledge base is not a one-time setup. Whenever your pricing, policies or product line changes, update the relevant document and re-upload it. Stale information is one of the most common causes of a bot giving wrong answers. Treat the knowledge base like a living FAQ page that needs the same maintenance discipline as your website.
Step by step
- Gather your source documentsCollect FAQs, pricing sheets and policy documents in one place before uploading.
- Open the agent's Knowledge Base sectionGo to the AI agent you want to train and find its Knowledge Base settings.
- Upload documents or URLsAdd PDFs, Word documents, text files or web page links.
- Wait for processingLet the platform index the content automatically before testing.
- Test with real customer questionsUse Preview / Try-it with the exact questions customers actually ask.
- Fix gaps in the source documentsEdit and re-upload any document that produced a vague or wrong answer.
- Review periodicallyRe-check and update the knowledge base whenever pricing or policy changes.
Example conversation
Train your AI agent free for 14 days
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
What file types can I upload to a chatbot's knowledge base?
Common document formats like PDF, Word documents and plain text, along with URLs to web pages, are typically supported for building a knowledge base. Good candidates include a clean FAQ, pricing sheet, policy pages and product spec sheets rather than a long, unstructured marketing brochure.
Do I need to format documents specially before uploading?
No special formatting is required, but documents with clear headings and short, self-contained paragraphs tend to produce more accurate answers than one long unstructured file.
How do I stop the AI from making up answers?
Instruct it explicitly in the system prompt to only answer from the knowledge base and to say when it's unsure, keep your documents accurate and non-conflicting, and set a lower temperature for factual agents.
How often should I update the knowledge base?
Whenever your pricing, policies or product details change. Treat it as a living document set, not a one-time upload, since stale information left behind after a price or policy change is one of the most common causes of a bot giving a confidently wrong answer.
Can one agent have multiple documents in its knowledge base?
Yes, you can upload several documents and URLs to the same agent's knowledge base, and it will search across all of them to find the most relevant answer.
What's the fastest way to find gaps in a knowledge base?
Run a batch of real, recent customer questions through Preview / Try-it in one sitting and note every vague or incorrect answer — that list tells you exactly which documents need clarifying or expanding.
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.