What the documents look like
Dealer rebate calculation happens after sales close. The basis for calculation sits across rebate policies, sales achievement statements, reconciliation statements and red-ink invoices. Contract ledgers record policy wording and settlement conditions; transaction materials record actual sales volume, payment collection and returns. These two groups must be aligned item by item before financial processing can proceed.
The calculation cycle usually runs monthly or quarterly. A single comparison session can involve contract clauses and transaction details spanning several months, so materials are long. This makes context and tokens a direct constraint at this step: the content length a model can reference in one request is finite, and it is necessary to judge which clauses and details must enter the comparison scope together.
Two or three facts drive the rest of this page. First, the comparison runs across separate documents rather than one file, so contract terms and actual amounts must appear together. Second, a rebate policy clause and a sales achievement statement can use different scopes, such as an annual agreement period and a single quarter, and the comparison must keep subject, period and field aligned. Third, the material set is long enough that tokens, not item counts, govern how much reference content fits.
Constraints at the context limits step
Model maximum context determines the content length a single request can carry. Reply limit controls output length only. These two are separate and cannot be substituted for each other. Knowledge base reference limit controls the maximum referenced content per search by tokens; when a single passage is long, it can be truncated, and key fields for the comparison may be missing.
Historical input changes the actual content sent for the same question under empty history versus history. The two conditions need separate checking. When a fixed reply format is used to carry comparison conclusions, field completeness and parse success need separate verification.
Hypothetical drill sample: a brand finance staff member checks one dealer rebate material set. Contract ledger text: "Party A signed an annual rebate agreement with Party B on 1 March 2025, agreeing that when the quarterly sales achievement rate is not lower than 90%, a rebate is calculated at 2% of the payment amount, with the agreement valid until 28 February 2026." Sales achievement statement text: "Party B's sales achievement rate for the second quarter of 2025 was 87%, and the payment amount was 1.2 million yuan." Draft sentence: "Party B met the rebate condition in the second quarter of 2025 and a rebate should be accrued at 2% of the payment amount." The question is whether this draft sentence is supported by the materials.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| AI model | Select a model that supports tool calling and whose maximum context covers the material length | Compare the maximum context shown in the configuration dialog with the total length of the sample materials |
| Reply limit | Enable and set a length that can hold the conclusion and its basis | Compare with the number of fields the sample draft sentence requires, and observe whether truncation occurs |
| Reference limit | Set by tokens, covering the contract clause and the key passage of the achievement statement | Compare with the length of a single passage, and check whether key fields are truncated |
| Minimum relevance | Set only when semantic retrieval or rerank is in effect | Observe whether the sample clause enters the referenced content; if missing, record the gap |
| Reply format | Plain text, json_schema not used for now | The sample conclusion has few fields, so fixed-structure parsing is not yet necessary |
| Memory turns | Keep the default | A single comparison does not depend on prior dialogue; check empty history and history separately |
Three common errors
- Treating the reply limit as the full context window. The reply limit restricts output length only; the content length a single request can reference is determined by maximum context. Check: input the full sample contract clause and achievement statement, compare the content range the model can actually reference with the total material length, and observe whether key clauses are missing because context is insufficient.
- Using a similarity setting as a substitute for permission or entity isolation. Minimum relevance filters low-relevance results only; it does not distinguish dealer entities or authorization scope. Check: in the referenced content, compare whether the contract subject and the achievement statement subject match; when the subject field is missing, record the evidence gap and do not judge authorization status from it.
- Performing joint association across libraries that use different vector models. Vectors generated by different vector models cannot be directly associated. Check: verify the vector model identifier used by each library, exclude libraries whose identifiers do not match, record the retrieval scope actually observed, and do not infer runtime results after association.
How to verify the setup
The following verification uses the hypothetical drill sample above and does not claim that any test was performed.
- Input the sample contract ledger text, the sales achievement statement text and the draft sentence. Compare "met the rebate condition" in the draft sentence with the contract term "quarterly sales achievement rate is not lower than 90%" and the "87%" stated in the achievement statement. When the same subject, same period and same field show mutually exclusive values, this is a contradiction. In the sample, 87% is below 90%, so the draft sentence is not supported and should be recorded as pending verification.
- Input the same comparison question under empty history and under history, and compare the difference in content actually sent to the model in the two runs. If under history the referenced content cites another dealer's material from prior dialogue, record that difference and check whether it affects the subject judgment for this sample.
- In the referenced content, check whether the contract validity period "until 28 February 2026" and the achievement statement period "second quarter of 2025" both appear. If the validity clause is truncated by the reference limit, record the evidence gap and do not fabricate an authorization or effective clause to obtain a definite conclusion.