What the documents look like
Fresh-food supplier claims pursue quality responsibility for goods on arrival. Store-level natural shrinkage is handled elsewhere, so the knowledge base for this category is organized around quality exception forms, photo evidence, deduction notices and claim reconciliation statements. Store service and internal training both draw on the same materials: how to fix evidence, how to calculate a deduction, and how to negotiate replenishment or compensation.
Materials update per batch and per month. A single batch's quality exception form and photo evidence are created on the acceptance day, while deduction notices and claim reconciliation statements are gathered monthly. Retrieval questions commonly carry a batch number, store name or exception type, so exact-identifier queries and natural-language questions appear side by side. The category difference matters because these documents record a supplier-facing quality liability, not store-level loss, and retrieval must return the right document type for the right stage.
A hypothetical paper exercise sample is used throughout this page. The submitted quality exception form reads: "Store: Chengbei; Batch: B-2405; Arrival date: 12 May; Exception: leafy greens wilted; Proposed handling: apply deduction." A photo evidence item is submitted with the description "outer carton label and wilted leaves from the same batch." The draft sentence under review is: "Chengbei store batch B-2405 leafy greens wilted, deduction completed per deduction notice." The draft asserts a completed deduction, yet the sample contains only a quality exception form and photo evidence. No deduction notice or reconciliation statement is present, so whether a deduction was executed is unsupported. A separate handover request reads: "Please synchronize the deduction result for this batch to the supplier business contact." That request is pending and does not constitute an executed action.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be associated together. Search mode is compared across semantic, full-text and hybrid retrieval against test questions; no single mode is assumed to suit all materials. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. Question optimization performs coreference resolution and question expansion. Image retrieval requires a corresponding image input and model capability, and re-ranking requires a model and must be enabled by configuration.
Retrieval returns cited content; it does not perform authorization checks and does not execute deductions or notifications. A pending request is not approval. An offered option is not execution. An unsigned termination is not a completed termination. The empty-reference output may have length zero, meaning the downstream path still proceeds even when nothing is retrieved.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Quality exception form base, photo evidence base | Both bases use the same vector model; verify the identifier before associating |
| Search mode | Hybrid retrieval | Compare recall differences using one identifier question set and one natural-language question set |
| Citation limit | 2000 tokens | Test content completeness with long and short exception forms and reconciliation statements, then compare against a smaller value |
| Minimum relevance | 0.3 | Label relevant and irrelevant samples first, then compare filtering against a value such as 0.5 |
| Question optimization | Enabled | Compare multi-turn follow-ups with single-turn explicit questions, checking whether the subject is altered |
| Result re-ranking | Enabled | When candidates are numerous, compare ordering before and after re-ranking and confirm the re-ranking model is available |
These starting values are hypothetical exercise points, not official defaults. Each is chosen so the comparison method can be stated and the observed result recorded.
Three common errors
- Treating retrieval recall as authorized execution. Input: the handover request "Please synchronize the deduction result for this batch to the supplier business contact." Compare this request against whether a deduction notice or reconciliation statement exists, and observe whether any executed record is present. With only a pending request, the result is not executed.
- Treating an unstated deduction completion as confirmed. Input: the draft sentence "deduction completed per deduction notice." Compare whether a deduction notice and reconciliation statement appear in the sample, and observe whether the materials support the wording. The sample lacks that field, so the result is unsupported.
- Associating knowledge bases across different vector models. Input: two knowledge bases with different identifiers. Compare their vector model identifiers and observe whether they match. If the identifiers differ, exclude the association and record the actual observation; do not run a cross-model retrieval comparison.
How to verify the setup
- Input the quality exception form text and the draft sentence from the hypothetical sample. Compare the deduction completion asserted in the draft against the fields the sample contains, and observe whether retrieval returns only the quality exception form and photo evidence. If the returned content does not include a deduction notice, the draft sentence is unsupported.
- Input the handover request "Please synchronize the deduction result for this batch to the supplier business contact." Compare the request status against any executed record, and observe whether an authorization or execution voucher exists. If only a pending request is present, record it as not executed and do not judge it completed.
- Input two knowledge base identifiers. Compare their vector model identifiers and observe whether they match. If they match, association is permitted; if they differ, exclude the association and record the actual observation. For empty-result handling, a clearly specified paper input
[]can be used to check the downstream design; this does not assert that a fictional rule retrieval would necessarily return[].
The checks describe observation and failure branches. They do not guarantee a first-hit result, full recall, or that an unrelated query returns empty. Where evidence is missing, it remains missing.