By Tina20 Nov,2026
When a vendor says you can “train a chatbot on your own data”, ask what happens technically. Many support products retrieve information from connected or uploaded sources rather than retrain the underlying model. That can be a practical approach, but it does not make every uploaded sentence true or current.Your preparation task is to give the assistant approved facts with clear scope, then verify that it uses them appropriately. Start with a small product set and known questions before importing every document the business has accumulated.
1. Ask how the product uses your data
Confirm which source types are supported, how updates become active and whether content can be restricted by shop, market or product. Ask how obsolete material is removed from active answering and what review controls are available.
Also ask whether the tool exposes supporting sources for an answer. If it does not, decide how reviewers will reliably check accuracy. Do not confuse a successful file import with successful answer quality.
| Supplier question | What you need to establish |
|---|---|
| Is this retrieval, fine-tuning or another method? | What a source update actually changes |
| Which records can the assistant access? | Scope and permissions |
| Does it access current stock or orders? | Whether a separate live connection exists |
| How are conflicting sources handled? | Your approval and exclusion process |
| Can a topic be paused? | A practical response to a serious error |
Create one dependable record for the product or variant that matters to the answer. Include a stable identifier, approved attributes, units, bundle contents and relevant exclusions. Keep marketing copy separate from factual claims where it obscures detail.
For an illustrative record: item code K-02, variant “with adapter”, cable included, adapter included, dimensions in centimetres, compatibility limited to the approved model list. A different “cable only” variant needs a separate scope rather than an exception hidden at the bottom of a long page.
Do not copy customer reviews into product facts as if every anecdotal claim were verified. Likewise, a sales slogan such as “works everywhere” is not an adequate compatibility specification.
Identify volatile fields. Inventory, prices and order status may need an authorised current source rather than a static upload. If the product cannot access that source, design a clarification or human check instead of implying real-time knowledge.
An FAQ should include the question, approved facts, applicability and transfer condition. “Do you accept returns?” is too broad if the answer depends on market, item and current order process.
Separate a general explanation from a specific approval. The assistant may explain where the buyer can review available return options while a person evaluates an exception. Do not train it to approve every request by placing an unconditional “yes” in a FAQ.
Rewrite ambiguous internal shorthand. A warehouse note saying “ship D+1 unless campaign” needs a verified buyer-facing explanation and defined exceptions. The assistant should not be expected to infer undocumented operating rules.
Choose the current approved policy version and record its effective date and owner. Archive old versions outside active answer scope. If two documents conflict, resolve the conflict before importing them.
Remove unnecessary customer identifiers, payment details and private staff information from examples. Use de-identified questions and invented test orders where possible. Data handling should follow the business's approved requirements and the supplier arrangement.
Do not upload an entire mailbox just because the product accepts documents. Past agent replies may contain mistakes, exceptional approvals and personal data. They are useful evidence for identifying questions, not automatically approved knowledge.
| Test input | Expected result | Why it matters |
|---|---|---|
| Known item and contents question | Correct contents for that variant | Basic source use |
| Similar item code | Correct distinction | Avoids catalogue mixing |
| Product missing | Clarification | Tests uncertainty |
| Expired campaign | No current offer claim | Tests validity |
| Unapproved compatibility | Human review or clarification | Prevents invented suitability |
| Refund exception | Transfer to authorised role | Tests authority boundary |
Include the expected source and pass condition. Without them, reviewers may approve an answer because it sounds plausible. Test informal wording and mixed-language questions representative of your buyers.
For important facts, test several phrasings. “What comes in the box?” and “adapter included?” should not receive incompatible answers for the same variant.
Start in a safe test or review mode where supported. Inspect a bounded live rollout closely. Classify errors as missing knowledge, contradictory knowledge, wrong scope, misunderstanding or connection failure.
Correct the cause. If the product reference never reaches the assistant, uploading more product descriptions may not help. If an old offer remains active, changing the tone instruction will not remove it.
Retest the original failure and nearby cases after every material change. Keep the source, change reason and approver in a simple log. Do not assume the system learns from an agent correction unless that specific controlled capability is available.
Duoke can be evaluated for approved product and service knowledge within its current supported scope. Confirm what Lite reads, how updates are managed and which questions remain human. The goal is dependable answers from maintained sources, not the largest possible import.

Email:[email protected]
Address:6/F MANULIFE PLACE, 348 KWUN TONG ROAD, KOWLOON, Hong Kong, 999077