What the documents look like
Store count and loss control records centre on book-to-physical differences and shrinkage. They do not handle supplier settlement. The working documents are count sheets, variance analyses, damage write-off approvals, and monitoring review records. Source data arrives from POS inventory, WMS, monitoring systems, and the SB/T count specification. The brief states that updates occur monthly or quarterly; where that cadence does not affect retrieval behaviour on this page, it is left aside.
For retrieval, this mix matters in two ways. Count sheets, variance analyses, write-off approvals, and monitoring review records carry precise identifiers such as shelf codes, SKUs, store identifiers, and period labels, alongside natural-language text describing variance causes and shrinkage. A single query may need to match a shelf code exactly and also match a phrase about why a variance occurred. The same corpus supports internal training explanations and store service wording checks, so retrieved passages are read by staff who compare a draft sentence against the underlying record.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Semantic retrieval relies on vector distance and suits natural-language questions. Full-text retrieval relies on keyword hits and suits precise queries such as codes, models, and proper nouns. Hybrid retrieval uses both recall paths and merges them through RRF. The reference cap is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimisation handles coreference resolution and question expansion. Image retrieval requires an image input and the relevant model capability. Reranking requires a rerank model and must be enabled in the settings. Retrieval itself does not perform permission or entity isolation.
Hypothetical paper exercise. A store count sheet for March 2024 records: "Category A shelf, closed-store count on 20 March, book quantity 120 units, physical quantity 114 units, variance 6 units, status is pending variance review." A damage write-off approval form for the same month records: "Category A shelf, 20 March variance 6 units, write-off approval status is pending approval." The draft sentence under review is: "The March count variance of 6 units for the Category A shelf has completed write-off approval." The scope of judgement is limited to the same store, the same shelf, the same period, and the same variance field. Both the count sheet and the write-off form show a pending status. The claim of completed write-off approval is not supported by either document. It remains unverified. This is not evidence that completion was negated.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Separate knowledge bases for count sheets, variance analyses, and write-off approvals | Check whether the vector model identifiers across the knowledge bases match; knowledge bases with mismatched identifiers are not associated together |
| Search mode | Hybrid retrieval | Compare code-type questions and natural-language questions against pure semantic and pure full-text retrieval |
| Reference cap | Hypothetical starting value of 2000 tokens | Compare whether content stays complete across count sheet fragments of different lengths, then adjust |
| Minimum relevance | Hypothetical starting value of 0.3 | Label relevant and irrelevant samples first, then compare how the filtered sample set changes |
| Question optimisation | Enabled for multi-turn follow-up questions | Compare the original question with the optimised question and check whether the count subject was altered |
| Result reranking | Enabled when candidates are numerous | Confirm a rerank model is available and enabled in the settings, then compare ordering before and after reranking |
Three common errors
- Treating a retrieval hit as a completed approval. Input the draft sentence together with the count sheet and write-off form text. Compare "pending variance review" and "pending approval" against "has completed write-off approval," and observe whether the status fields agree. When the statuses disagree, the draft sentence is not supported.
- Associating knowledge bases that use different vector models. Check the vector model identifiers of each knowledge base, exclude those that do not match, and record what is actually observed. Do not run retrieval across mismatched models to compare missed recalls.
- Treating minimum relevance as a permission or entity filter. Input a cross-store question, compare the returned content against the access scope, and observe whether records outside that scope are recalled on similarity alone. Retrieval does not perform authorisation.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the status fields in the count sheet and the write-off form against the wording of the draft sentence, and observe whether the draft sentence is supported by the source text. The expected outcome is not supported and pending verification.
- Input the count sheet fragment and the write-off form fragment. Compare the same store, the same shelf, the same period, and the same variance field, and observe whether the two values are mutually exclusive. The expected outcome is that they are not mutually exclusive; both are pending.
- Input an empty reference result. Check the downstream chain design against an array length of 0, and observe whether the downstream branch handles the empty result. The expected outcome is that the chain runs through. Do not claim that retrieval will necessarily return an empty array.