How to Build an AI Knowledge Base for Ecommerce Customer Support

By Tina17 Nov,2026


An AI knowledge base is the approved information an assistant can use to answer questions. For an ecommerce shop, the hard part is rarely collecting enough text. It is making sure the text applies to the correct product, variant, shop and date.
Uploading an old campaign guide alongside a current product sheet can create a confident but incorrect answer. The assistant may find a plausible sentence while missing the condition that made it true. Build the knowledge around answerable tasks and maintained sources, not around the number of files imported.

1. Start with the questions buyers actually ask

Review recent conversations across products, stores and shifts. Identify the information agents repeatedly search for: included accessories, measurements, compatibility, dispatch arrangements, return routes and promotion conditions.
Choose a bounded initial set. A knowledge base covering the most common verified product questions can be useful before it covers every unusual after-sales scenario. Keep questions requiring live order information or discretionary approval in a separate category.
For each question, ask what evidence would let an agent answer correctly. If nobody can locate a dependable source, the problem is a knowledge gap, not a prompt-writing problem. Assign a person to resolve it before automatic answering.

2. Separate four information types

Information type Suitable content Important boundary
Product knowledge Attributes, variants, contents and approved compatibility Similar items may differ
Service knowledge Staffed hours, standard processes and approved policies Scope by shop and market
Campaign knowledge Offers, eligibility and validity Expire or restrict obsolete terms
Current operational data Stock, tracking and order status Needs an authorised current source where supported

Do not store a snapshot of inventory in a static FAQ and call it live stock. Do not treat a general return explanation as approval for a specific buyer's claim. The knowledge base can explain a process while a person retains the decision.
Keep platform rules distinguishable from the shop's own practices. A warehouse preference cannot override the buyer's applicable platform process. Link to the current authoritative source and give someone responsibility for reviewing changes.

3. Use a compact record with explicit scope

Field What to record
Topic or question The buyer task the record supports
Approved answer facts Verified information, not only polished wording
Applies to Shop, market, product and variant
Source Authoritative listing, policy or approved internal record
Validity Effective date and expiry where relevant
Owner Person responsible for correction
Transfer condition When the assistant should ask or hand off

For an illustrative bundle record, identify the exact variant and included items. Add the condition that a missing product reference requires clarification. Avoid a universal answer such as “all packages include a cable” when the catalogue contains exceptions.
Keep the record readable by humans. If an agent cannot tell when it applies, an assistant will be difficult to audit. A long paragraph mixing three shops and several exceptions is usually harder to maintain than separate scoped records.

4. Resolve conflicts before importing

Create a source hierarchy appropriate to the business. For example, the approved current product record may take precedence over an archived campaign note. The hierarchy must be decided by the responsible team, not improvised by the assistant.
Look for duplicated questions with different answers, old refund wording, expired offers and measurements in inconsistent units. Preserve useful history outside the active answer scope when needed, but do not leave obsolete content available as if it were current.
When two sources disagree, record the disagreement and assign an owner. Temporarily exclude or restrict the topic if a wrong answer would be material. A clarification or human check is better than asking the model to choose whichever wording sounds more convincing.
Do not assume a system automatically learns every corrected agent reply. Unless that feature is explicitly available and controlled, update the approved source through the actual maintenance process.

5. Keep language versions tied to the same facts

For shops serving the Philippines, Malaysia and Singapore, buyers may use English, local languages or mixed wording. The facts should remain consistent even when the phrasing changes.
Preserve product codes, measurements, amounts and conditions. Have a capable reviewer check important translations, especially exclusions, compatibility and service commitments. A natural-sounding translation can still reverse the meaning of “not included”.
Test informal questions rather than relying only on formal FAQ wording. The assistant should recognise a common product question without inventing a fact that the approved source never supplied.

6. Build a test set with expected answers

Use examples representing the actual catalogue and service workload. Remove unnecessary customer identifiers. For each test, record the expected source, required clarification or expected transfer.
Test category Example Pass condition
Direct fact Dimensions of known variant Correct value and unit
Ambiguous reference “Does this include it?” Requests the missing context
Similar products Two nearly identical model names Does not mix their attributes
Expired campaign Previous offer mentioned Does not present it as current
Missing knowledge Unapproved compatibility claim Clarifies or transfers
Authority boundary Refund exception No invented approval

Test the same fact in different wording and languages relevant to the shop. Include conversations where the buyer changes product midway. A one-question demonstration cannot show whether the assistant maintains the correct context.
Score factual correctness, applicability and uncertainty handling separately from tone. Do not let elegant wording compensate for the wrong item specification.

7. Launch with human review

Start with a topic or shop scope the team can inspect. If supported, use review-before-send or suggested answers during evaluation. Otherwise use an approved restricted test path and a controlled live rollout with close monitoring.
Assign a reviewer who understands the product facts. Record failures as knowledge missing, source conflict, retrieval or scope problem, misunderstanding, or workflow failure. These categories point to different corrections.
For example, adding a source will not repair a connector that fails to supply the product reference. Rewording a prompt will not resolve two contradictory packing lists. Fix the actual cause and retest the affected questions.

8. Maintain knowledge when the business changes

Build review into existing work. A catalogue update should trigger an attribute check. A campaign ending should remove or restrict its active offer. A policy change should update the approved record and tests that depend on it.
Trigger Owner action Verification
Variant changed Update contents or attributes Retest variant questions
Offer expires Retire active offer wording Ask an old-offer question
Repeated unanswered enquiry Investigate the gap Add approved facts and a test
Wrong answer found Assess severity and correct cause Retest original and related cases
Staff or shop change Recheck scope and ownership Confirm who maintains each record

Keep a change log with the date, reason, source and approver. This makes an unexpected answer easier to investigate without blaming the agent who happened to receive the complaint.

9. Measure usefulness rather than library size

Track whether common questions have approved coverage, whether answers are accurate and whether unresolved issues reach people. A larger knowledge base is not necessarily better if it contains duplicates and obsolete conditions.
Review repeated contact and human corrections. An apparently successful answer followed by the same unanswered question suggests a gap in substance or context. Buyer silence alone is not reliable evidence of resolution.
Duoke can be evaluated for using approved ecommerce knowledge in routine answering across supported channels. Confirm current Lite functions, product scope and maintenance controls. Do not assume live data access or automatic gap learning merely because the assistant can read product information.
A useful knowledge base is one the business can keep true. Start with that standard, then expand as the catalogue and service process can support it.

Further reading


Android Version
iOS Version
International Edition

Contact Us

Email:[email protected]

Address:6/F MANULIFE PLACE, 348 KWUN TONG ROAD, KOWLOON, Hong Kong, 999077

Service and Support

Facebook

WeChat Official Account

Rednote

TikTok

FairCloud Technology (HK) Limited