What the documents look like
Near-expiry markdown work follows the shelf-stock life cycle. It is a different operating shape from item-by-item weighing and sorting. The main data sources named in the brief are POS stock age, store inspection, ERP inventory, and quality inspection records. The daily documents are near-expiry lists, markdown approvals, damage write-off forms, and promotional price tags.
Document comparison for contracts and transaction materials places the subject, object, status, and validity scope from these documents onto the same field for checking. The retrieval step only recalls candidate references from materials already stored in a knowledge base. Whether a markdown, write-off, or handover is authorized is decided downstream by approval and business systems, not by retrieval. The brief gives a data update frequency of daily or per shift. Retrieval configuration does not by itself create batch synchronization or real-time triggers.
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 with different vector models cannot be associated together. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process, not a standalone switch. The minimum relevance score takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold and does not perform permission or entity isolation. The reference limit is measured in tokens and controls the maximum amount of content cited per search. Question optimization performs coreference resolution and question expansion, and it adds one model call. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled by configuration. Reference content is output as an array, and its length can be zero.
Hypothetical drill sample: A store inspection record for the shift of 10 March 2025 states: "Near-expiry list number LX-0310, object is refrigerated yogurt batch A, shelf life to 12 March 2025, status is pending markdown approval." A draft markdown approval form of the same date states: "Agree to sell batch A at a 30 percent discount, approver signature pending." A draft sentence reads: "Batch A markdown has been approved and can be executed at a 30 percent discount." To be judged: whether the draft sentence and the two materials share the same subject, the same object, and the same status.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Comparison method and reason |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base holding near-expiry lists and markdown approvals; vector model identifiers are consistent | Check vector model identifiers first, exclude inconsistent ones, do not run cross-model retrieval |
| Search mode | Start by comparing semantic and full-text recall with hybrid retrieval | Number LX-0310 suits full-text hits; status descriptions suit semantic recall |
| Reference limit | Hypothetical starting point set to a token amount sufficient to hold both source materials | Compare with the source text length and observe whether references are truncated |
| Minimum relevance score | Hypothetical starting point at a middle value, effective only under semantic retrieval or reranking | Compare against labeled relevant and irrelevant samples and observe filtering behavior |
| Question optimization | Off first, then on for short questions for comparison | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Not enabled yet; compare later when candidates are numerous | Reranking requires a model and must be enabled by configuration; score range is 0-1 |
Three common errors
- Reading a pending markdown approval status as already approved. Input is the inspection record and approval draft above; compare the status field. The draft says "approver signature pending," the draft sentence says "has been approved," and the same object carries mutually exclusive values in the same field, so the draft sentence is unsupported. Executable check: compare the status field character by character and observe whether "signature pending" and "has been approved" appear together.
- Treating retrieval recall as authorization execution. Input is the reference array returned by the search node; compare the reference content with the markdown approval form and observe whether a signed approval conclusion appears in the references. When only a draft appears, the authorization status remains pending verification. Executable check: check the signature column of the source document in the references and record missing items.
- Forcing association of knowledge bases with different vector models. Input is the vector model identifiers of two knowledge bases; compare whether the identifiers match and observe whether any mismatch exists. On finding a mismatch, exclude it and do not continue comparing recall results. Executable check: list each base identifier and mark conforming and non-conforming ones.
How to verify the setup
- For the draft sentence in the hypothetical sample, input the retrieval question "LX-0310 batch A markdown approval status," compare the recalled references with the inspection record and approval draft source text, and observe whether the references contain the "signature pending" wording. If the references contain only the draft and no signed conclusion, the draft sentence is judged unsupported.
- For the same object, input the question "batch A markdown approver," compare the recalled content with the signature column of the approval form, and observe whether a signed approver appears. When none appears, record it as an evidence gap and do not judge it as a denial or conflict.
- For the empty-result branch, use an explicit paper input of an empty array to check the downstream design and observe whether the downstream continues along the path where the reference length is zero. Do not claim that a fictional rule makes retrieval always return an empty array. All three checks describe expected observations only and do not claim they were actually tested.