What the documents look like
Retail checkout and loyalty material comparison in this category depends on four document types: promotion price tags, membership points rules, checkout daily reports, and anomalous refund slips. These materials originate from the POS system, membership CRM, payment platform, and price tag system, with daily or per-shift update frequency. Store checkout supervisors, CRM operations staff, and finance auditors need to reconcile these materials within the same comparison task, confirming whether promotion prices, points entitlements, and refund records remain consistent. The subject, object, status, and effective scope recorded in the materials form the basis for judgment. A promotion price tag states the applicable store and effective period. A membership points rule states the points calculation method and applicable membership tier. A checkout daily report states the shift transaction summary. An anomalous refund slip states the refund reason and authorization status. These fields are distributed across different documents, and comparison requires confirming the value relationship of the same subject, same period, and same field item by item.
In the context and token step, the relevant constraints involve the content length the model can reference, the knowledge base citation limit, the historical input range, and structured output handling. The model maximum context is checked against actual model capability, and the reply limit controls single-response length; these are distinct concepts. The knowledge base citation limit uses tokens as its unit, with each search citing at most n tokens of content. A long single passage may be truncated, and the actual cited content needs to be checked against the question sample. Historical input affects retrieval behavior in continuous conversation, and the same question may produce different input under empty history and with history. Structured output requires verifying field completeness and successful parsing.
Hypothetical drill sample: A store promotion price tag states in its original text: "From June 1 to June 30, 2025, Class A members purchasing designated cooking oil enjoy a 20% discount, limited to 2 barrels per card." The membership points rule states in its original text: "Class A member consumption points accrue at 1 point per 1 yuan actually paid; promotional items do not participate in points." The checkout daily report states in its original text: "June 15 shift, Class A member Zhang purchased 2 barrels of designated cooking oil, actual payment 96 yuan, points increase 0 yuan." The anomalous refund slip states in its original text: "June 15, Zhang applied to refund 1 barrel of cooking oil, refund amount 24 yuan, refund reason: duplicate purchase, authorization status is pending review." Draft sentence to be judged: "Zhang purchased cooking oil on June 15 and received 96 points; after the refund the points have been deducted." This draft sentence is inconsistent with the points rule original text "promotional items do not participate in points" and the checkout daily report "points increase 0 yuan." The refund slip authorization status is pending review, which cannot be treated as a completed refund or points deduction.
Constraints at the context limits step
The model maximum context and the reply limit serve different functions. The maximum context governs the content length a single request can reference. The reply limit governs the length of a single answer. Setting a reply limit does not expand the referenceable content, and a large maximum context does not guarantee the reply will include every field.
The knowledge base citation limit is measured in tokens. When a single passage of material is long, the cited content may be truncated, and the comparison conclusion may lack the field needed for judgment. The actual cited content needs to be checked against the question sample rather than assumed to cover all four documents.
Historical input affects retrieval scope. The same question under empty history and with history may produce different retrieval input, and the cited content may differ accordingly. Verification needs to compare the actual cited content under both conditions rather than assume the first hit or full recall.
Structured output handling requires verifying field completeness and successful parsing. A fixed reply format such as json_schema needs its output fields checked against the comparison conclusion requirements, and parsing success needs to be observed rather than inferred from text length.
The minimum relevance score is a filter threshold that takes effect only under semantic retrieval or result reranking. It does not perform permission or entity isolation and does not exclude unrelated stores or unauthorized records. Retrieval is not authorization.
Thinking budget is used for the model thinking process and does not serve as the citation token limit.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason for selection and comparison method |
|---|---|---|
| AI model | Select a model supporting the required context length | Compare the actual model maximum context against the total length of the four materials to confirm whether the price tag, points rule, daily report, and refund slip full text can be accommodated |
| Reply limit | Set to a length sufficient to output the comparison conclusion | Compare against the number of basis fields required by the draft sentence to confirm the reply is not truncated |
| Knowledge base citation limit | Set by token to a value covering key passages of the four materials | Compare against single passage length and the question sample to observe whether cited content is truncated |
| Minimum relevance score | Set according to the effective conditions of semantic retrieval or reranking | This value is a filter threshold and does not carry permission or entity isolation functions |
| Reply format | Select json_schema when fixed fields are required | Verify comparison conclusion field completeness and parsing success |
| Thinking effort | Select a level according to question complexity | Thinking budget is used for the model thinking process and not for the citation token limit |
Three common errors
- Treating the reply limit as the full context window. The reply limit controls single-response length, while the maximum context controls the content length a single request can reference; these are different. Check method: Input the four materials in full, compare the content the model can actually reference against the reply limit value, and observe whether the reply is truncated due to the limit.
- Treating the minimum relevance score as permission or entity filtering. The minimum relevance score is a filter threshold effective only under semantic retrieval or result reranking, and it does not exclude unrelated businesses or unauthorized records. Check method: Input a query containing an unrelated store, observe whether returned results are filtered only due to low relevance, and confirm no permission filtering logic is present.
- Treating a pending-review refund slip as a completed refund. The refund slip authorization status is pending review, and it cannot be inferred that the refund has been executed or points have been deducted. Check method: Input the refund slip original text and the draft sentence, compare the authorization status field against the draft sentence wording, and observe whether the draft sentence writes pending review as completed.
How to verify the setup
- Input the four material original texts and the draft sentence from the hypothetical sample, compare the draft sentence "received 96 points" against the points rule "promotional items do not participate in points" and the checkout daily report "points increase 0 yuan," and observe whether the comparison conclusion identifies the sentence as inconsistent with the original text. Expected result: the draft sentence is not supported.
- Input the refund slip original text from the hypothetical sample, compare the authorization status "pending review" against the draft sentence "after the refund the points have been deducted," and observe whether the comparison conclusion retains pending review as a to-be-verified status. Expected result: neither refund execution nor points deduction is supported.
- Input the same question under empty history and with history, compare whether the retrieval input changes due to different history turns, and observe whether the cited content differs. Expected result: historical input affects retrieval scope, and the actual cited content needs to be checked; the first hit or full recall is not guaranteed.
For an empty-result downstream check, a paper input of [] can be used to verify the downstream design, without claiming that a hypothetical rule retrieval necessarily returns []. Across different vector models, libraries cannot be jointly associated; first check the identifier and exclude non-conforming ones, and do not execute cross-model retrieval to compare missed recall. The optional paths listed in the overall document flow diagram do not mean every request executes all of them: image retrieval requires the corresponding image input and model capability, and reranking requires the model and its enabled setting.