What the documents look like
Creator special livestream sales cooperation centers on external creator fulfillment and commission settlement. Brand business teams, MCN agencies, and creator managers work daily with creator cooperation agreements, commission setting sheets, sample shipment sheets, and livestream performance reports. These materials update per session or per week, and a single livestream session may produce multiple commission setting sheets and performance reports. The document comparison task places key fields from the same session, same creator, and same product side by side across different materials to judge whether values agree, whether they fall within a valid range, and whether they sit at the same authorization stage.
The typical data sources are the selected alliance, the Juliang Xingtu platform, creator historical data, and store transaction details. Errors in this work lead to commission settlement disputes, fake traffic, and damaged brand reputation. The three most time-consuming steps are screening creators, negotiating commission and slot fees, and reconciling after-sales commission deductions.
Constraints at the context limits step
Comparison requires sending relevant passages from multiple materials into the model context at the same time. The maximum context of a model determines the content length a single request can reference. The reply limit only controls output length, and the two must not be conflated. The knowledge base reference limit controls the content amount referenced per search by token, so a long single passage may be truncated, and problem samples must be used to check whether the fields needed for judgment are retained. In multi-turn follow-up scenarios, the memory turns setting affects the prior context range sent to the model. The same question under empty history and under history-bearing conditions produces different inputs, and the comparison conclusion may change. When a fixed reply format carries comparison results, field completeness and parseability must be verified.
Hypothetical exercise sample. Material text for checking: The creator cooperation agreement states "Party A commissions Party B to promote product A in the special livestream on June 18, 2025, commission rate 15%, settlement based on the transaction amount confirmed by the livestream performance report." The commission setting sheet states "Product A, June 18, 2025 special session, commission rate 15%, status is pending review." The sample shipment sheet states "Product A sample, 2 units, shipped June 10, 2025, receipt status: signed." The livestream performance report states "June 18, 2025 special session, product A transaction amount 80,000 yuan." Draft sentence to be judged: "The commission for this session has been settled at 15%." Judgment points: the commission rates in the agreement and the commission setting sheet agree, both 15%; the commission setting sheet status is pending review, with no indication of approval; no settlement completion record appears in the materials. The draft sentence's "has been settled" is not supported by the materials and should be recorded as pending verification.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| AI model | Select a model whose maximum context covers the needed length | Check whether the model maximum context covers the sum of relevant passages from the four materials, compared against total material length |
| Memory turns | Retain a small number of history turns | Comparison tasks are mostly single-pass material checks; compare against the same question under empty history |
| Reply limit | Retain enough length to list each field comparison conclusion | Compare against the number of fields required by the draft sentence, confirming no early truncation |
| Knowledge base reference limit | Set by token, covering the longest single material passage | Compare against the token amount of the longest single passage, confirming judgment fields are not truncated |
| Reply format | Select json_schema when structured carrying is needed | Verify returned fields are complete and parseable, compared against field-by-field check results |
| Minimum relevance | Set according to the conditions under which semantic retrieval or rerank takes effect | Compare against recall results, confirming low-relevance fragments are filtered; this does not perform permission filtering |
Three common errors
- Treating "pending review" on the commission setting sheet as approved. Input is the commission setting sheet text; compare the status field value against the wording "approved"; the observable result is that the status field contains no approval wording, and this should be recorded as pending verification. Judging settlement as complete directly amounts to treating pending review as approved.
- Treating the commission rate agreed in the agreement as an already settled amount. Input is the agreement and performance report text; compare the rate in the agreement against the transaction amount in the report; the observable result is that these are different fields, and the settlement amount requires separate calculation and confirmation, not direct derivation from the rate.
- Treating a retrieval hit as authorization confirmation. Input is the agreement passage returned by knowledge base retrieval; compare the retrieval result against the authorization status field; the observable result is that retrieval returns only text content and does not return approval status, so authorization judgment requires checking approval records separately.
How to verify the setup
- Input the four material texts and the draft sentence from the hypothetical sample; compare the draft sentence "has been settled" against the commission setting sheet status and the performance report amount field in the materials; the observable result is that no settlement completion record appears in the materials, and the draft sentence should be marked as pending verification. If the model outputs "has been settled" without marking the evidence gap, the configuration does not meet the comparison requirement.
- Input the same question under empty history and under history-bearing conditions; compare the context range sent to the model in each case; the observable result is that the memory turns setting changes the prior context range. If both inputs are identical while the setting has changed, check whether memory turns took effect.
- Input a single material passage exceeding the reference limit; compare the content sent to the model after truncation against the key fields of the original text; the observable result is that the fields needed for judgment are still retained. If key fields are truncated, adjust the reference limit or send in segments. For empty-result handling, a paper input of
[]can be used to check downstream design, without claiming that retrieval necessarily returns an empty array.