What the documents look like
For order fulfilment and after-sales answers in a brand member private-domain operation, the working materials are the member tag table, the outreach plan, the coupon rules, and the privacy authorisation record. These materials come from the CRM system, enterprise WeChat, POS purchase records, and mini-program behaviour data, with update frequencies of daily or weekly. A traceable after-sales answer aligns each cited order status, coupon rule, and authorisation scope against the corresponding material line by line. The member tag table determines who may be contacted, the coupon rules determine redemption conditions, and the privacy authorisation record determines the usable scope. When any of these is missing, the answer can only remain in a pending-verification state.
The materials are not uniform in form. A member tag table carries a member identifier, a tag value, an owning store, and an authorisation scope. A coupon rule carries a coupon identifier, an applicable order status, and a validity end date. A privacy authorisation record carries a member identifier, an authorised purpose, and an authorisation period. An after-sales question that cites an order status, a coupon condition, or an authorisation range must be checked against the same subject, the same field, and the same period in these records.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. A zero-length output still passes the output link through. The citation upper limit is controlled in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Hybrid retrieval combines results through RRF. Reranking requires a rerank model and must be enabled according to the setting. Retrieval is not authorisation, and a similarity score does not isolate business entities or filter unauthorised records.
The following is a hypothetical paper exercise sample, created for this page and not drawn from any real record.
Material sentence one, from the member tag table: "Member identifier A-001, tag value pending verification, owning store S-01, authorisation scope limited to order inquiry."
Material sentence two, from the coupon rules: "Coupon C-88, applicable order status completed, valid until 30 June 2025."
Material sentence three, from the privacy authorisation record: "A-001 authorised purpose after-sales answers, authorisation period 1 January 2025 to 31 December 2025."
Draft sentence to be judged: "Member A-001's order may use coupon C-88 to process a refund."
The comparison uses the same subject A-001, the same field order status, and the same period 2025. The materials do not state a refund authorisation. The draft sentence is therefore not supported by the materials, and the evidence gap must be recorded. The draft sentence is not contradicted, because no material states an exclusive opposite value for the same field.
Configuration and implementation checks
The configuration below is tied to the hypothetical sample. It records documented settings and implementation checks, not database internals.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Member private-domain order after-sales library | Confirm the selected library and material version; only knowledge bases with the same vector model may be associated together |
| Search mode | Hybrid retrieval | Compare hit results against semantic retrieval and full-text retrieval |
| Citation upper limit | Hypothetical starting value 800 tokens | Compare against the sample material length to confirm the citation is complete |
| Minimum relevance | Hypothetical starting value 0.5 | Compare against the rerank score to filter low-relevance results |
| Question optimisation | Enabled | Compare retrieval results for follow-up questions against the disabled state |
| Citation display check | Line-by-line comparison with source text | Compare against the three sample material sentences; answer fluency does not replace source evidence |
The citation upper limit is expressed in tokens, the same unit the documentation uses for that setting. The minimum relevance value is a 0-1 number and acts as a filter threshold, not a final ordering rule. Question optimisation adds a model call before retrieval, which can improve follow-up retrieval and also adds latency.
Three common errors
- Treating a retrieval result as authorisation. The input is the retrieval output, the comparison object is the privacy authorisation record, and the observable result is whether the answer exceeds the authorised purpose. An answer that exceeds the authorised purpose is not supported by the record.
- Treating a pending review as an approval. The input is the review status field, the comparison object is an explicit value in the same field, and the observable result is whether an approval record exists. When no approval record is stated, the status remains pending verification.
- Associating knowledge bases that use different vector models. The input is the knowledge base identifier, the comparison object is the vector model identifier, and the observable result is whether the identifiers match. When they do not match, the knowledge base is excluded and the actual observation is recorded. No cross-model retrieval is performed to compare missed recalls.
How to verify the setup
- Input the three sample material sentences and the draft sentence. Compare the subject A-001, the field order status, and the period 2025. Observe whether the draft sentence is supported by the materials. When it is not supported, record the evidence gap. The expected outcome is that the refund statement remains unsupported, because no material states a refund authorisation.
- Input an explicit paper empty array as the citation output. Compare whether the downstream design accepts an empty citation array. Observe whether the workflow continues. The expected outcome is that the downstream link proceeds, because the documented output length may be zero. This check does not claim that a fictitious rule search must return an empty array.
- Input two knowledge base identifiers that carry different vector model identifiers. Compare whether the identifiers match. Observe whether the non-matching knowledge base is excluded. The expected outcome is that the non-matching knowledge base is excluded and the observation is recorded. This check does not run a cross-model retrieval to compare missed recalls.
For each check, the material must be available within the target retrieval scope with access rights, or supplied as an explicit manual input. When the material is missing, the gap is recorded rather than filled with an assumed value. A pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination.