A customer asks whether a refund is available. Your agent finds a current FAQ that says 30 days, an old PDF that says 14, and a campaign page with no date. The problem is not a lack of chatbot training sources. It is that nobody decided which source is allowed to answer.
Source reliability matters more than source volume. Ten conflicting pages create a worse customer experience than one maintained answer. Use this rule throughout your setup: choose each source based on the question it must answer, assign an owner, and define the event that triggers a review.
Start with the answer, not the source
Chatbot training sources are the files, website content, pasted text, and exact Q&A pairs an AI agent uses to answer customers. Each source type is useful. None should be treated as a general-purpose bucket for every fact your business has ever written down.
Start with the questions your customers ask. A refund rule needs precise wording. A product explanation needs context. A customer asking for an order update may need an API call rather than a paragraph from your knowledge base.
That distinction changes what you upload. You are not trying to give the agent the largest possible library. You are building a small set of approved answers that can be checked, maintained, and retired when the business changes.
Rank your sources by answer reliability
Use this order when deciding which content should answer a question:
- Exact Q&A pairs for fixed wording. Use them for refund rules, eligibility requirements, legal language, security statements, and other answers where a paraphrase can create trouble. A Q&A pair gives the agent a narrow answer instead of asking it to interpret a long document.
- Approved product documentation and help-center pages. These work well when customers need an explanation, steps, or related context. Check that the page reflects the current product before adding it.
- Selected website pages. Website crawling gives you broad coverage. It can also pull in navigation pages, old announcements, campaign copy, and duplicate pages. Review the pages first instead of crawling the whole site by default.
- Uploaded internal files. PDF, DOCX, and TXT files are useful for material that is not public, such as internal guidance or approved operating instructions. AssistLoop accepts uploaded files up to 42MB each. Train your agent on your own data for the supported source workflow.
- Pasted text for small changes. Use pasted text when a short, controlled addition needs to go live without editing a full document. It is a good place for one new answer while a larger policy or help-center page is being updated.
The ranking is about the type of answer, not a permanent score for each format. A current policy PDF may be more trustworthy than an old FAQ page. A short Q&A pair is still the better source when the wording must stay exact.
For each source, record the question it is meant to answer. If you cannot name that question, the source probably does not belong in the first training set.
Do not combine sources that disagree
Suppose a current refund Q&A says customers have 30 days to request a refund. An old policy PDF says 14 days. Adding both sources and hoping the agent chooses correctly is a bad process. The agent has been given two plausible answers and no reliable way to know which policy your team enforces.
Resolve the conflict before training:
- Mark one page or policy as canonical.
- Remove old copies from the active knowledge base.
- Update the canonical source when the rule changes.
- Use an exact Q&A pair when the final wording must be precise.
Keep customer-facing product facts separate from internal notes, draft copy, and temporary launch material. A launch document can explain what the team plans to release. It should not answer a customer after the feature, price, or policy has changed.
A source register makes this work visible. Keep four fields for every active source:
| Field | What to record |
|---|---|
| Source name | The page, file, pasted entry, or Q&A pair |
| Owner | The person responsible for its accuracy |
| Last review date | When someone last checked the source against the live policy or product |
| Allowed questions | The customer questions this source may answer |
When two sources disagree, the register gives you a place to resolve the dispute. It also stops a source from becoming a permanent part of the agent simply because nobody remembers why it was added.
This approach follows the same basic concern described in the NIST AI Risk Management Framework: identify and manage risks instead of treating an AI system as a one-time installation. For website content and broader publishing practices, Google’s guidance on creating helpful, reliable, people-first content is a useful standard too. If a page is outdated for a human reader, it is a poor training source for an AI agent.
Build a maintenance loop for stale answers
Training is not finished when the first source is uploaded. Give each source a review trigger:
- Product releases trigger a review of product documentation.
- Policy changes trigger a review of exact Q&A pairs.
- Site edits trigger a review of crawled pages.
- Retired plans, features, prices, and policies trigger removal from the active source set.
Use your conversation logs and analytics to find the work that needs attention. Look for unanswered questions, incorrect answers, and questions that repeatedly lead to human handoff. A question that appears several times is a better training task than a guess about what customers might ask next.
A missing answer does not always require a new document. If the gap is narrow, add one exact Q&A pair. For example, if customers keep asking whether a specific plan includes a feature, add that answer directly after confirming the current plan details. Do not upload an entire sales deck to fix one missing sentence.
Keep old versions out of the active knowledge base. A retired plan can look like a valid plan to an agent if its page is still available in the training set. The same applies to old prices, discontinued features, and replaced policies.
AssistLoop retrains automatically when training content changes. That reduces the mechanical work. It does not decide whether a new page is accurate, whether an old PDF should be removed, or whether two sources disagree. Your maintenance task is choosing the correct source and keeping it current.
Know when the agent should stop answering
Some questions should go to a person. Route billing disputes, account-specific issues, sensitive requests, and questions with conflicting source material to your team instead of forcing an answer.
A clear handoff rule is simple: if the agent cannot support an answer from an approved source, it should collect the available context and pass the full conversation to the team. That is safer than a confident answer built from a weak match.
With human handoff, the team receives the full thread on paid plans, along with country, browser, and device details. The customer does not need to repeat the conversation from the beginning.
Handoff patterns also tell you where source quality needs work. A cluster of handoffs around one question can mean the answer is missing. It can also mean the answer exists but is too vague, split across conflicting pages, or reserved for a human because it requires account access.
Do not treat every handoff as a failure. A billing dispute should reach a person. A missing shipping policy may call for a new Q&A pair. The useful question is why the handoff happened and what action, if any, belongs in the training set.
Put the playbook into AssistLoop
Start with a small source set. Create an agent, then add the highest-authority material first:
- Add exact Q&A pairs for fixed wording, such as one canonical refund policy.
- Add approved files for internal guidance that customers are allowed to receive.
- Add selected website links for public product information.
- Review the first conversations before adding more material.
In those first conversations, check three things. Did the agent use the right policy? Did it say when the available information was insufficient? Did it hand off when the question required a person?
Expose the agent through the widget only after those checks pass. Widget customization covers the customer-facing setup, including the name, greeting, avatar, colors, theme, and suggested replies. The presentation matters, but it comes after source quality. A polished widget still gives a bad experience if the policy behind it is wrong.
Add an Agent Action when the source workflow cannot finish the job. Checking order status through an API, booking a meeting, and capturing a lead are examples of work that needs an action rather than another paragraph in the knowledge base. Keep the source answer and the action separate: the source explains the process, while the action handles the live request.
Plan the test with AssistLoop pricing and message credits. Each user message and each AI response counts as one credit, and credits do not roll over.
Last verified: August 2026
AssistLoop is the wrong fit if you need the agent to answer from constantly changing account data without connecting an approved API or action. Training content can explain a process. It cannot safely invent a current account balance, order status, or billing result.
Your next step is small: create your AI agent, load one canonical policy and a small set of real Q&A pairs, then inspect the conversations before expanding coverage.
FAQ
What are chatbot training sources?
Chatbot training sources are the files, website pages, pasted text, and exact Q&A pairs an AI agent uses to answer customers. The right source depends on the question: use exact Q&A pairs for fixed wording, approved documents for context, and selected website pages for broad public coverage.
Which source is best for a chatbot policy answer?
Exact Q&A pairs are usually the best choice when an answer must use precise wording, such as a refund rule, eligibility requirement, or legal statement. They limit the agent to a focused answer instead of asking it to interpret a long document.
How should I handle conflicting chatbot training sources?
Do not leave both sources active and hope the agent picks the newer one. Mark one source as canonical, remove or update the conflicting copy, and use an exact Q&A pair when the final wording must stay precise.
How often should chatbot training sources be reviewed?
Review sources when the related product, policy, or website content changes. Conversation logs can also show when an answer is missing, incorrect, or repeatedly handed to a person, giving you a specific source to update.
Written by