What the documents look like
Fresh-food supplier claims pursue responsibility for quality at arrival and do not handle natural in-store shrinkage. That boundary shapes copy drafting: proposed sentences must correspond item by item with the quality exception note, photo evidence, deduction notice and claim reconciliation statement, and promotional wording absent from those materials cannot be written in. The main data comes from the quality inspection system, WMS acceptance, supplier contracts and financial reconciliation, with updates per batch and per month. On this page the drafting context is assembled from field fragments of those documents, and token consumption concentrates in the quality exception description, the deduction basis and the reconciliation conclusion.
Constraints at the context limits step
The length of content a model can reference in a single request is set by model capability, while the reply limit only caps output length; the two cannot substitute for each other. The knowledge-base reference limit is controlled by tokens, so when a single material passage is long the referenced content may be truncated, and the fragments actually included need to be checked against the question sample. A larger memory turn count more easily carries earlier conversation into the request and also raises consumption. When a fixed reply format is used, field completeness and successful parsing must be verified separately; text length is not evidence of successful downstream handling.
Hypothetical drill sample: the material text reads "Batch A-0312 spinach dated 12 March 2025; the acceptance record states that the leaf wilting ratio exceeds the upper limit agreed in Appendix 2 of the contract; quality inspection number QZ-0312-07; deduction notice KD-0312-02 has listed a deduction amount; the claim reconciliation statement has not been signed." The draft sentence to be judged is "This batch of spinach suffered total cargo loss due to supplier responsibility, and the supplier has confirmed compensation." The objects to compare are: the material states only that the wilting ratio exceeded the limit and that the deduction notice has listed an amount; it does not state total cargo loss and does not state that the supplier confirmed compensation. The draft sentence exceeds the range supported by the material and should be rewritten to match the stated facts, or marked as pending verification.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison object |
|---|---|---|
| Reply limit | Length sufficient to generate a complete claim explanation | Compare with a short reply of one sentence to confirm the basis and handling action are not truncated |
| Reference limit (token) | Covers a single quality exception note and deduction notice | Compare with the length of a single material passage to observe whether drafting fields are truncated |
| Memory turns | Retain a small number of history turns | Compare with an empty-history input to rule out drafting drift caused by context differences |
| Reply format | Select structured output when fixed fields are required | Compare with default text output to verify field completeness and successful parsing |
| Minimum relevance | Takes effect only in semantic retrieval or with reranking | Compare with results when the value is off to observe whether low-relevance fragments are filtered |
These values are hypothetical starting points for this trial, not official defaults. Retrieval is a recall process and does not carry permission or entity isolation; libraries across different vector models cannot be associated together, so the library identifier must be checked first and non-matching ones excluded, with the actual observation recorded.
Three common errors
- Writing a promotional sentence that the material does not state directly into the copy. The draft sentence says "total cargo loss" and "has confirmed compensation," while the material states only that the wilting ratio exceeded the limit and that the deduction notice has listed an amount. These are different fields and different fact states, so the sentence is unsupported.
- Treating the reference limit as the reply limit. The reference limit controls the token volume brought into the model, and the reply limit controls output length; mixing them up can truncate the drafting basis while the output still looks complete.
- Treating a pending-review or unsigned status as completed. The claim reconciliation statement has not been signed, and that is a different fact from a completed signature and from confirmed compensation; a definite conclusion cannot be written on that basis.
How to verify the setup
- Input the hypothetical sample material text and the draft sentence, and compare the subject, object and status in the draft sentence against the fields stated in the material. When the draft sentence contains "total cargo loss" or "has confirmed compensation" that the material does not state, it is judged unsupported and must be rewritten or marked as pending verification.
- Input the same question under two conditions, empty history and a small number of retained history turns, and compare the context content included in each request. When the drafting basis is the same in both, the history input introduced no extra drift; when a difference in basis appears, record the differing fields and check the memory turn setting.
- Use an explicit paper input of an empty result and check the downstream handling design. Observe whether the downstream handles it as an empty reference, without claiming that retrieval will necessarily return an empty result. When structured output is used, compare the returned fields against the required field list; a missing field or a parsing failure counts as not passing.
The above is paper checking only and does not mean a test, deployment or business action has actually been carried out. Authorization, effective clauses and compensation confirmation that the materials do not state remain gaps, and no clauses are fabricated to obtain a definite answer.