A chatbot knowledge base works when it contains clear, current answers to real customer questions, with a named owner for each source and a safe response when no answer exists. Before a widget goes live, review what the crawler found, remove stale or conflicting material, define escalation, then test the questions customers actually ask. Uploading every file in the company drive is a fast way to make a customer-facing bot less dependable.

An AI knowledge base chatbot retrieves company information to help answer a visitor. It does not make its source material trustworthy by itself. The practical job is to decide which pages, files, policies and FAQs are allowed to answer public questions, then keep that set current. SketricGen supports knowledge bases built from crawled websites, uploaded files, pasted text and Q&A pairs, with crawled material available as editable Markdown. (Source: SketricGen Knowledge Bases)

Key takeaways

  • Give every public source an owner, review date and the customer question it is meant to answer.
  • Keep public product, support, policy and FAQ content. Remove duplicate, stale, internal and boilerplate material before activation.
  • Treat two sources that answer the same question differently as a release blocker, not a minor cleanup task.
  • Tell the chatbot what to do when its sources do not answer: clarify, route to a person or decline the claim.
  • Retest after a source, policy, product or pricing update. A successful first launch says little about the next change.

What belongs in a chatbot knowledge base

Start with material a customer should be able to hear from a support representative: current product pages, setup documentation, account and cancellation policies, shipping or service terms, troubleshooting articles, and carefully written FAQs. Each source should answer a recognizable question such as "Can I cancel mid-cycle?" or "Which plan includes this feature?" A page that only announces a campaign is usually a weak source, even if it is public.

Website crawling is often a sensible first pass because it captures the language already presented to visitors. It still needs review. SketricGen lists website crawling, file uploads, pasted text and Q&A pairs as inputs, and says crawled pages can be edited before an agent uses them. (Source: SketricGen Knowledge Bases)

Inference: organize sources around customer decisions, not department folders. A billing policy and an invoice FAQ can live in different systems, but the chatbot needs one approved answer for the billing question. This is how a website chatbot knowledge base becomes useful rather than merely large.

Once the approved source list is clear, create a SketricGen Brand Agent from your website and test it against the customer questions in your inventory. Start with a narrow public scope. Add more sources only after the first set passes review.

What should stay out

Leave out outdated pricing pages, duplicate policies, internal playbooks, employee-only documents, generic boilerplate pages and claims the company cannot support. A page can be publicly reachable and still be a bad chatbot source. Old launch posts, archived terms and regional pages written for another market often create contradictions around prices, eligibility or product availability.

Internal documents deserve special care. They may include confidential pricing, escalation notes, unapproved roadmap language or instructions that make sense only to employees. Zendesk notes that unauthenticated customers can use only public restricted content when permissions are configured. That is a platform-specific behavior, but it captures the broader rule: public agents need deliberately public source scope. (Source: Zendesk Knowledge Sources)

Inference: exclude navigation, footer repetition, cookie notices, login pages and thin landing pages unless they contain the only approved answer to a customer question. SketricGen specifically calls out repeated navigation/footer text and pages that should be removed after a crawl. A smaller approved set is easier to review and retest. (Source: SketricGen Knowledge Bases)

Prepare a chatbot trained on website content, PDFs and FAQs

To train a chatbot on your own data, use one preparation pass per source type. For website pages, crawl only the public paths that serve support, product or policy questions. Review the extracted text, remove repeated layout text, fix missing context, then identify the owner. SketricGen describes this crawl review as editable Markdown, which makes cleanup part of the source workflow instead of a separate copywriting project. (Source: SketricGen Knowledge Bases)

For PDFs, prefer a text-based original over a scanned brochure. Give the file a descriptive name and make sure headings, tables and policy exceptions exist as selectable text. If a plan comparison relies on a chart image, write the relevant rules in text too. Intercom advises teams to include text explanations beside screenshots or other rich content because a bot may not interpret media in the same way a customer does. (Source: Intercom Guide)

FAQs should pair one question with a complete, standalone answer. Restate the question in the answer where context could otherwise be lost. Structured content such as a price table, eligibility matrix or supported-country list needs a single owner and a version or review date. OpenAI similarly separates reference material from instructions: the source gives the agent facts, while configuration tells it how to behave. (Source: OpenAI GPT Knowledge)

If you are building a public, company-trained website agent, build it with SketricGen Brand Agents. Brand Agents are the customer-facing product path for website pages, support docs, product pages, pricing, policies and FAQs. (Source: SketricGen Knowledge Bases)

How to handle contradictions and stale information

Do not ask the chatbot to "choose the newest" source when two policies disagree. Make the content owner resolve the conflict before the agent can answer. Keep the rejected source out of the active set or rewrite both sources to point to the same canonical policy. This matters most for pricing, eligibility, security commitments, refunds and regulated claims.

External-source freshness is not always live. Zendesk says its connected external sources generally reflect the last sync, usually every 24 hours, while help-center searches use current help-center content. The timing is specific to Zendesk, but it is a useful reminder to record how each platform refreshes a source before publishing a changed policy. (Source: Zendesk Knowledge Sources)

Inference: define a conflict rule with three states. Green means one current approved source. Amber means a source changed and needs retesting. Red means two active sources disagree, so the agent must not answer from either until an owner chooses the canonical wording. A short source inventory makes this visible before a visitor finds the mismatch.

Unknown-answer behavior and escalation paths

The bot needs behavior rules as well as source material. Reference content tells it what it may say. Instructions decide what happens when retrieval is weak, the request is sensitive or the source set conflicts. OpenAI makes the same separation in its GPT guidance: use knowledge for reference material and instructions for rules or behavior. (Source: OpenAI GPT Knowledge)

Set a plain escalation path. First, ask a clarifying question if the visitor's goal is unclear. Next, answer only when the approved source covers the question. When the answer is missing, state that the chatbot cannot confirm it and offer the right human route, such as support, sales, billing or a security contact. Intercom describes knowledge-base agents as able to clarify and escalate when the available information is insufficient. (Source: Intercom Guide)

Inference: do not let the agent estimate a price, promise an integration, interpret a contract or invent a policy exception. Those are handoff triggers. Record the destination, operating hours and information to collect in the escalation rule, then test the handoff with an actual ticket or form submission.

Source inventory template

Use this copy-ready template before loading or activating a source. Add one row for every page, file, FAQ collection or structured dataset. Leave the retest status as Not tested until the expected question produces the approved answer.

Source URL or file | Owner | Last reviewed date | Customer question covered | Conflict risk | Retest status

| | | | |

Use the same release rules for every row:

Conflict riskMeaningRelease action
GreenOne current, approved answer existsKeep the source active and retest after changes
AmberThe source changed or needs owner reviewKeep it out of scope until review and retest finish
RedTwo active sources disagreeBlock the affected answer until the owner selects the canonical source

Inference: make the inventory the release record. A red row stops activation. An amber row can remain out of scope until its owner finishes the update. "Passed" should mean a named question was tested after the latest source change, not merely that the file uploaded successfully.

Chatbot knowledge base best practices: launch and retest

Before launch, confirm the agent can cite or link its approved help content where the product allows it, and verify it does not use excluded documents. Check at least one easy question, one ambiguous question, one answer that requires a policy exception, one unavailable-answer question and one human handoff. Capture the expected source or expected abstention for every test.

After launch, review unresolved conversations and search terms. Zendesk recommends monitoring resolved and unresolved conversations and the terms customers ask, while SketricGen recommends retesting after knowledge updates. NIST's AI Risk Management Framework frames evaluation as part of the lifecycle, not a one-time approval. (Sources: Zendesk Knowledge Sources, SketricGen Knowledge Bases, NIST AI RMF)

Retest whenever pricing, terms, eligibility, product availability, onboarding steps or escalation contacts change. Also retest when a crawler runs again, a file is replaced or a conversation exposes an unsupported answer. The checklist is ordinary operational work. It is also the best defense against a polished chatbot repeating an old truth.

When the inventory is clean and the test set has expected answers, launch a SketricGen Brand Agent on a limited set of pages first. Review unresolved questions before expanding it across the whole site.

FAQ

What is the knowledge base in a chatbot?

A chatbot knowledge base is the approved set of pages, files, FAQs and structured information the chatbot can retrieve to answer customer questions. It should contain current public information and clear ownership, rather than every document a company has. The chatbot's behavior rules should remain separate from that source material. (Sources: SketricGen Knowledge Bases, OpenAI GPT Knowledge)

How do you build a knowledge base for a chatbot?

Audit customer questions first, select the current public sources that answer them, clean the extracted text, assign owners and record review dates. Resolve duplicate or conflicting answers before activation. Then test direct questions, unclear questions and questions with no supported answer. SketricGen Brand Agents supports the public website-agent workflow described here. (Sources: Intercom Guide, SketricGen Knowledge Bases)

Can a chatbot trained on website content answer from PDFs too?

Often, yes, if the platform supports file uploads and the PDF contains readable text. SketricGen lists PDF among its supported ingestion types. A PDF with vital information locked in an image or an unreviewed old version can still be a poor source, so test the precise question it is expected to answer. (Source: SketricGen Knowledge Bases)

How often should I update a chatbot knowledge base?

Update it whenever a customer-facing fact changes, especially pricing, policies, eligibility, availability or product steps. Also account for the platform's sync behavior. Zendesk says external sources can reflect the last sync rather than live content, and SketricGen recommends retesting after a knowledge update. (Sources: Zendesk Knowledge Sources, SketricGen Knowledge Bases)

Related coverage

References