What the documents look like
Seller ratings maintenance draws on platform experience scores, service indicators, after-sales reviews and customer service systems. The working documents are negative-review handling records, experience score breakdowns, service diagnostic reports and rectification ledgers. When drafting product and marketing copy, these materials supply post-fulfilment consumer feedback facts, and any promotional sentence in the draft needs to correspond to the source wording in those materials.
For this page the relevant material properties are length and field structure. A single passage may exceed the reference limit, so key fields can be trimmed away. Earlier input also shapes how the model reads the current question, so the same question can produce different answers with empty history and with history present. When a fixed reply format is used, field completeness and successful parsing need separate verification, because acceptable text length does not mean the structure is correct.
Constraints at the context limits step
Three limits operate at different layers. The maximum context of a model determines how much content a single request can reference. The reply cap controls how long one answer may be. The knowledge-base reference limit is set in tokens and controls how much retrieved content is carried into the answer. Treating any one of these as another produces drafting errors.
Hypothetical paper exercise sample. Source material reads: "A shop's March 2025 experience score breakdown shows a logistics experience score of 4.2, a service experience score of 4.5 and a product experience score of 4.1; the negative-review handling record states that on 12 March one negative review completed contact and supplementary evidence collection, with handling status pending review; the service diagnostic report concludes that logistics timeliness is the main deduction item; the rectification ledger records that on 15 March a rectification plan was submitted and no review result has been received." The draft sentence to be assessed is: "This shop has excellent logistics experience, all negative reviews have been fully handled, and the rectification plan has been approved." Three points require verification: whether a logistics experience score of 4.2 supports the wording "excellent", whether the status "pending review" equals "fully handled", and whether "submitted" equals "approved".
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | Choose a model whose maximum context covers the material | Compare the model's maximum context against the estimated total tokens of the material so the material is not truncated |
| Reply cap | Set a length that can hold the basis and the handling action | Compare against the number of fields the draft sentence needs; too short a cap truncates the basis |
| Reference limit | Set in tokens to cover the longest single passage | Compare against the token count of each passage so key fields are not trimmed |
| Memory rounds | Retain a small number of history rounds | Compare inputs for the same question with empty history and with history in a multi-turn follow-up |
| Reply format | Select json_schema when a fixed structure is required | Compare field completeness and parsing outcome; acceptable text length does not mean correct structure |
| Minimum relevance | Set according to the conditions under which semantic retrieval or rerank applies | This value is a filtering threshold and does not perform permission or entity isolation |
Three common errors
- Treating the reply cap as the full context window. A model's maximum context and the reply cap are two independent settings. The reply cap only controls the length of one answer and does not determine how much material can be referenced. Compare the model's maximum context value against the estimated total tokens of the material.
- Reading the status "pending review" as "fully handled". When the status field in the negative-review handling record is pending review, that fact is not complete. Compare the source status field against the wording in the draft sentence; when the status field does not state completion, record the wording as unsupported.
- Reading "submitted" in the rectification ledger as "approved". The submission action and the review result are different facts. Compare the submission date and the review result field in the ledger; when the review result is absent, keep the evidence gap open.
How to verify the setup
- Enter the source material and the draft sentence from the hypothetical exercise sample, then compare the phrase "excellent logistics experience" against the logistics experience score of 4.2 in the experience score breakdown. When the material does not state a criterion for "excellent", the wording is unsupported and the gap is recorded.
- Enter the negative-review handling record source text and compare the status field "pending review" against "fully handled" in the draft sentence. These are different values of the same field; when the status is pending review, the draft wording conflicts with the material and is judged not to hold.
- Enter the rectification ledger source text and compare the submission date against the review result field. When the ledger records only the submission action and no review result, the draft phrase "has been approved" remains to be verified and cannot be judged as approved. If a fixed reply format is used, separately compare whether returned fields are complete and whether parsing succeeds. For empty-result handling, a paper input of an empty array can be used to check the downstream design; this does not assert that a fabricated rule search will necessarily return an empty array.