Knowledge base chatbot: turn store information into useful answers

Start with questions that block a purchase

A customer rarely asks for a knowledge base. They ask whether a product fits, which service includes installation or what happens if they need to return an item. Organise the source material around those questions.

Collect a small sample of recurring, anonymised questions. Group them by the decision they support: choosing, buying, delivery or after-purchase help. For each, identify the person who owns the correct answer. Avoid uploading every company document before you know what the customer needs.

Choose the right source for each fact

Published policies and product guides can explain relatively stable facts. Current price, available stock and private order status need a source that can answer for the current request. A document describing last month’s catalogue is not a live inventory connection.

Write an answer record a person can maintain

Use a consistent record for each important topic. This is a content template, not a required Noodle Seed file format.

• Customer question: the wording people actually use.
• Answer: the fact and its conditions in plain language.
• Applies to: product, region, plan or customer type.
• Source: the approved page or system.
• Owner and review trigger: who updates it and what makes it change.
• If uncertain: the clarification or contact route to offer.

For example, “Can I wash this cover?” needs the fabric, care instructions and any exception. “Easy to clean” is marketing language, not a washing instruction. Use the manufacturer’s supported care guidance rather than a plausible guess.

Resolve conflicts before adding more documents

Duplicate answers make maintenance harder. Choose one source for each policy, remove retired versions from the published set and make exceptions visible beside the general rule. Keep internal notes and customer-specific records out of public sources.

When two approved pages disagree, decide which owner must resolve the conflict. The assistant should acknowledge a missing fact or route the question to a person rather than manufacture a compromise. A larger knowledge collection is useful only if it improves the answers.

Test answers and source changes

Create a small evaluation set with the expected answer and source for each question. Include a paraphrase, a question with a missing detail, an unsupported request and a conflict between old and new information. Then change one test source and repeat the affected cases.

Record whether the answer is accurate, whether its conditions are preserved and whether the next step is useful. Keep unanswered questions separate from incorrect answers. Do not count a confident response as a pass when its source is missing.

Use the customer experience evaluation toolkit to structure observations. These checks are a process for your team, not a claim that a particular supplier has passed them.

Connect knowledge to the next customer action

Once answers are reliable, choose the next useful action: show the right product, ask for a missing requirement or continue to the store’s checkout. Knowledge alone does not grant access to a customer account or permission to change an order.

Noodle Seed’s Shopify implementation guide illustrates a path from store information to product selection and checkout handoff. For a different business, agree the sources, supported actions and customer journey before implementation.

Read the ecommerce guide for a buying-flow checklist, or compare Noodle Seed and Chatbase when choosing an approach.