What the documents look like
Near-expiry markdown work concerns the remaining sellable time of shelf inventory rather than per-item weighing at an order station. The four recurring documents are near-expiry lists, markdown approvals, damage write-off forms, and promotional price tags. These materials describe batches, time limits, and disposal actions: how many days remain before a stated expiry date, whether a batch has entered a markdown window, which approval step a markdown request has reached, and whether a write-off form and a price tag refer to the same batch. Source records arrive from POS stock age, store inspections, ERP inventory, and quality inspection records. These four sources do not describe the same batch with identical fields, so a batch number, a date, and a store may appear in different positions or formats across them.
Vector retrieval in this setting maps natural-language questions from store staff, such as how many days a batch can still be sold or which batches can no longer be placed on the shelf, onto fragments of those materials. Retrieval returns cited content only. It does not change inventory state, and it does not constitute a markdown approval or an isolation action.
Constraints at the vector indexes step
The smallest search unit is the library. Collections exist for classification management and do not affect search results. Libraries associated with one search node must use the same vector model, so model selection questions require checking the actual model and input type before choosing libraries; libraries with non-matching model identifiers are excluded from the same retrieval node.
One original data block can correspond to multiple vectors, and the indexes field records their vector IDs. A hit on any one vector recalls that original data. Markdown approvals and write-off forms often carry a batch number, a date, a store, and a disposition conclusion. Building separate indexes for the batch number and the conclusion makes different phrasings more likely to reach the same block than compressing the whole passage into one vector. Shorter, more focused index content narrows retrieval coverage accordingly, which suits scenarios with strict answer boundaries.
Image retrieval depends on whether an image description or image vector was generated at ingestion time. Supplying an image only at the query side does not compensate for an index missing at the ingestion side. Hybrid retrieval merges multiple result paths through RRF, which is a result fusion process rather than a separate switch. Minimum relevance is a filter threshold that takes effect only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation, and a similarity score cannot be used to exclude records the requester is not authorized to view.
Configuration and implementation checks
The table below lists checkable items around one paper exercise sample for this page. The sample is paper material and does not indicate that any real business action occurred.
Hypothetical exercise sample: a store inspection record states "2026-03-10 inspection, batch A low-temperature yogurt, 42 units, 3 days to the labeled expiry date, markdown approval submitted, approval status pending review." A draft sentence for the same day reads "Batch A has completed markdown approval and may be placed on the shelf with a markdown price tag." A separate handover request reads "Please retrieve the effective markdown approval record for batch A from the library and place it on the shelf accordingly."
| Check item | Value or practice in this page's sample | Comparison object and criterion |
|---|---|---|
| Vector model identifier | Record the vector model identifier actually used by the selected library | Compare character by character with identifiers of other libraries in the retrieval node; exclude non-matching ones |
| Index content | Build separate indexes for batch number and approval status | Compare whether the index text contains subject, object, and status |
| Index and original data | Check the original data block the indexes field points to | One original data block may map to multiple vectors; after a hit, retrieve the source text for comparison |
| Image index | No image input in the sample, not involved | Without image input this path does not execute; draw no conclusion |
| Citation limit | Set to 2000 tokens as this exercise's starting value | Chosen to accommodate several batch fragments; compare with a smaller value for citation completeness |
| Minimum relevance | Set to 0.3 as this exercise's starting value | Chosen to observe filter strength; compare with a higher value for recall count |
The citation limit and minimum relevance are starting values for this exercise, not official defaults. The comparison method fixes the same question set and changes one value at a time, observing changes in cited source text.
Three common errors
- Treating pending review as approved. The sample source text reads "approval status pending review," while the draft sentence writes "has completed markdown approval." The input is the inspection record and the draft sentence; the compared field is approval status; the two values are mutually exclusive on the same subject, batch, and date, so the draft sentence is not supported by the source text.
- Treating a retrieval hit as authorization. The handover request asks to "retrieve the effective record and place it on the shelf accordingly." Retrieval returns cited content only. Shelf placement is an inventory disposal action that requires separately confirmed approval and execution steps; the retrieval result does not constitute that authorization.
- Associating libraries with different vector models. When two libraries have non-matching model identifiers, the correct practice is to check the identifiers and exclude the non-matching one, recording the actual observation. Do not infer recall behavior after joint association.
How to verify the setup
The following three items address this page's sample and are paper checks, not executed tests.
- Input the inspection record source text and the draft sentence, and compare the two status fields "pending review" and "has completed markdown approval." Under the same subject, batch, and date the values are mutually exclusive; the observable result is that the draft sentence conflicts with the source text, so it is judged unsupported.
- Input the handover request and compare the action "place on the shelf" with the retrieval output content. The retrieval output is a citation array whose length can be 0; the observable result is that the output contains no authorization or execution record, so the request is judged to lack an authorization basis and the evidence gap is retained.
- Input the vector model identifiers of the two libraries and compare whether they match character by character. The observable result is that a non-matching library is excluded and no cross-model joint retrieval is executed; only matching identifiers may enter the same retrieval node, and the actual observation is recorded.
For empty-result handling, a downstream design can be checked with an explicitly paper-input empty array, without inferring that a certain class of question necessarily returns empty. After model or index adjustments, compare cited source text using the same question set, and do not promise effects based on vector scores alone.