What the documents look like
Price and display audits take place during ordinary store operations, and the materials centre on continued execution of a standard. The supervision system records store visit results, the POS price list reflects the price actually charged in the store, store cameras and visit photographs supply display evidence, the headquarters promotion calendar states the price and display requirements for the current period, and rectification notices carry the follow-up loop. These materials are updated weekly and monthly, and a single comparison usually covers a small number of stores and a limited set of products within one promotion period.
For the contract and transaction material comparison task, the context and token constraints come from the shape of the material itself. Visit photographs and price tag lists often arrive as images or short text, while the promotion calendar and rectification notices contain dates, store numbers, product names and price fields. One comparison has to hold the standard text, the store execution evidence and the draft under review inside the same model context at the same time. The reference limit is controlled in tokens, and a single long passage can be truncated.
Constraints at the context limits step
The context window is checked against the actual model capability, and the reply limit only restricts output length. The two are separate quantities. The knowledge base reference limit is measured in tokens; in a mixed knowledge base the passage lengths differ greatly, and control by token is more stable than control by item count. Historical input affects retrieval, so the same question produces different input with empty history and with history, and each case needs separate checking.
Image retrieval requires the corresponding image input and model capability, and reranking requires a model and must be enabled in the settings. The paths drawn in the overall document flow diagram do not mean that every request executes all of them. Semantic recall and relevance scores do not carry permission or entity isolation, and the requirement that a retrieval library and its vector model match does not extend to external systems. Libraries built on different vector models cannot be associated together; the identifier is checked first and non-matching libraries are excluded.
Hypothetical exercise sample: a supervisor in the operations department of a chain brand checks price tag execution at Store A during the June 2025 promotion period. The source text supplied for comparison is: "Headquarters promotion calendar, 2025-06-01 to 2025-06-30, product M standard retail price 19.9 yuan, member price 17.9 yuan." "Store A visit photograph shows that on 2025-06-15 the shelf price tag reads 21.9 yuan, with no member price marking." The draft sentence under review is: "During the June 2025 promotion period, Store A's price tag price matches the headquarters standard and the display meets the requirements." This draft sentence must be compared against the two source passages above, with Store A as the subject, June 2025 as the period, and the price tag price of product M as the field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Model context | Check against the stated value of the selected model | The standard text, store evidence and draft sentence must fit together; compare with total material length |
| Reply limit | Keep enough to express the basis and the handling action | Compare with the number of fields the draft sentence needs; too short truncates the stated basis |
| Knowledge base reference limit | Set a token value from the length of a single passage | Compare with the single-passage length of the price tag list and promotion calendar to see whether truncation occurs |
| Historical input | Check once with empty history and once with history | Compare the retrieval results for the same question under the two inputs |
| Reply format | Select json_schema when fixed fields are required | Compare with downstream parsing needs; verify complete fields and successful parsing |
| Minimum relevance | Check against the conditions under which semantic retrieval or reranking applies | Compare with candidate scores to see whether the target material survives filtering |
These values are hypothetical starting points for this exercise, not official defaults. The reference limit and the thinking budget are different concepts: the thinking budget controls model reasoning cost, while the reference limit controls the amount of retrieved content.
Three common errors
- Treating the reply limit as the whole context window. The reply limit only restricts output length; the amount of content a model can reference in a single request is set by the context. During checking, compare the total material length with the stated context value and observe whether the material is truncated.
- Using the minimum relevance score to exclude unrelated businesses or unauthorized records. Relevance is a filter threshold and does not carry permission or entity isolation. During checking, compare candidate content before and after filtering and confirm that filtering only affects score ordering, not the accessible range.
- Treating retrieved results as authorization or proof of execution. Retrieval returning material does not mean the business action was approved or performed. During checking, compare the status fields in the material; pending review, quoted and signed correspond to different facts, and authorization that is not stated remains a gap.
How to verify the setup
- Using the hypothetical exercise sample above, enter the promotion calendar source text and the visit photograph description, then compare the price field in the draft sentence with the standard price of 19.9 yuan in the source. The draft states the price matches, while the source shows 21.9 yuan; the same field carries mutually exclusive values, which is judged a contradiction. If the source does not state the price field, record it as unsupported.
- Using the same hypothetical sample, enter the same question once with empty history and once with history, and compare the sets of material returned by the two retrievals. If the history-bearing input recalls material outside the promotion period, record the difference and check whether the history content introduced an unrelated period.
- Using the explicit paper input of an empty result
[], check the downstream handling design. Observe whether the downstream outputs an empty reference or a notice that no material was found, without asserting that a fictional rule retrieval necessarily returns an empty array. If retrieval returns material, compare the material subject with the requesting subject, exclude non-matching items and record the actual observation.