What the documents look like
Near-expiry markdown handling follows the shelf-life of store inventory. It does not weigh and sort individual order lines. The main data sources are POS stock-age records, store inspections, ERP inventory, and quality inspection records. Everyday documents include near-expiry lists, markdown approvals, damage reports, and promotional price tags. These materials together determine whether a batch may stay on sale, at what price, and when it moves to damage reporting. Retrieval questions usually concern batch identifiers, date ranges, approval status, and supplier material. The documents are produced on a daily or shift-level rhythm. That rhythm describes when records appear; it does not imply batch synchronization, real-time triggers, or transaction linkage.
Constraints at the retrieval step
A knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. A 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. Retrieval can run as semantic search, full-text search, or hybrid search. Hybrid search merges multiple result paths through RRF. ReRank requires a model and must be enabled in settings. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reordering is used. Question optimization performs coreference resolution and question expansion, and it adds one model call. Cited content is returned as an array whose length can be zero, and the output path still completes when nothing is found. Retrieval and relevance scoring do not perform permission or entity isolation.
Hypothetical paper exercise sample: a fresh store's shift inspection record for 12 March 2026 reads: "12 March morning shift inspection, Area A cold cabinet batch 20260310-A, shelf life until 15 March, outer packaging intact, held in the pending markdown area." The markdown approval record for the same batch reads: "Batch 20260310-A markdown application submitted, status pending review, no approver signature seen." A draft sentence reads: "Batch 20260310-A markdown has been approved and may be placed on sale with a markdown price tag." The question to decide is whether that draft sentence matches the two materials on the same subject, the same period, and the same field value.
Configuration and implementation checks
The table below lists settings or implementation checks that the official documents describe, tied to this exercise sample. The values are hypothetical starting points for this exercise, used for comparison against another value. They are not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Include only the store inspection and approval libraries that use the same vector model | Check the vector model identifier first; exclude libraries whose identifiers differ, and do not run cross-model retrieval comparisons |
| Search mode | Hybrid search | Batch identifiers are exact queries; run semantic and full-text paths and compare the candidate differences |
| Citation limit | Hypothetically set to 800 tokens | Test content completeness with near-expiry lists of different lengths, then compare against 400 |
| Minimum relevance | Hypothetically set to 0.4 | Label relevant and irrelevant samples first, then observe how filtering changes the sample set |
| Result reordering | Enable when a model is available | Compare candidate order before and after enabling; the score range is 0-1 |
| Question optimization | Enable for multi-turn follow-ups or short questions | Compare the original question with the optimized question and check whether the subject was altered |
Three common errors
- Treating a retrieval hit as authorization. In the sample, the markdown approval status is pending review. Finding that record only shows the material exists. The draft sentence "markdown has been approved" lacks signature evidence. Pending review does not show that approval was given, and it does not show that operating permission is absent.
- Treating RRF in hybrid search as a switch that can be turned on. RRF is the result-fusion process inside hybrid search. The configuration entry should record the search mode and the reordering enablement state, not an RRF switch.
- Treating minimum relevance as a permission or entity filter. That value is a 0-1 filtering threshold and takes effect only under semantic search or reordering. It cannot be used to exclude unauthorized records or unrelated businesses.
How to verify the setup
- Input the two source materials and the draft sentence above, then compare subject, period, and field values. The draft sentence claims approval; the approval record field says pending review with no signature. The same field holds mutually exclusive values, so the claim is unsupported and the evidence gap must be recorded.
- Input the plain-text query "batch 20260310-A" and observe whether the cited content array contains the inspection and approval source text for that batch. If an empty array is returned, check the downstream handling design against a clearly stated paper input of []. Do not claim that a fictional rule makes retrieval always return [].
- Input an exact query containing only the batch identifier, run it once as hybrid search and once as pure semantic search, and compare the candidate sets. Observe whether full-text recall adds identifier hits. Do not guarantee a first-position hit or full recall. Record a failure branch as a missing candidate and keep it in a pending-verification state.