What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so the materials handled here are the store's own transaction and operating records rather than franchise contract files. The main data sources are the POS system, the membership CRM, the ERP inventory system, and store inspections. Day-to-day documents include daily sales reports, shift schedules, display photos, and inventory transfer orders. These materials sit in several systems with differing formats and field definitions, so before any comparison the retrieval step has to confirm that the same entity, the same period, and the same field are actually comparable. Updates arrive daily and weekly, and retrieval reflects the state of the material at the time it was indexed; it does not carry a real-time synchronization duty.
A second practical feature is that a single store generates both numeric records and narrative records about the same events. A daily sales report states amounts, a transfer order states quantities and dates, and a display photo carries no structured field at all. Retrieval can return all of these together, and the comparison task then depends on whether the returned passages share a field that can be checked against each other.
Constraints at the retrieval step
A knowledge base is made up of libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. One data item can map to several vectors, and a hit on any one of them recalls that item, with the vector score taken as the highest. Semantic search relies on vector distance, full-text search relies on keyword hits, hybrid search merges multiple result paths through RRF, and ReRank reorders candidate text and produces a 0–1 score. Minimum relevance takes effect only under semantic search or when result reordering is used; it is a filtering threshold and does not perform permission or entity isolation. When several knowledge bases are associated, they must use the same vector model, and cross-model association is not allowed; identifiers should be checked and non-matching bases excluded rather than running a cross-model search to compare missed recalls. The reference limit is measured in tokens, and when the reference output has length 0 the downstream path still runs.
Hypothetical paper exercise sample: the original text of a daily sales report for a brand-owned store for March 2025 reads "1 March to 31 March, category A goods sales total RMB 120,000." The original text of an inventory transfer order for the same period reads "5 March, 200 units of category A goods transferred in from the regional warehouse." The sentence proposed for judgment reads: "The store's March category A sales of RMB 120,000 and the 200 units transferred in can be mutually corroborated in quantity." Both materials share the same entity and the same period, but the fields are sales amount and transferred-in units, which are not mutually exclusive values of the same field. The judgment should be recorded as unsupported, and same-field material must be supplied before further checking.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge base | Associate only the store transaction materials library | That library uses one vector model and matches the material scope under review |
| Search mode | Hybrid search | The materials contain both identifiers and natural-language description; compare against pure semantic and pure full-text results |
| Reference limit | Set to a token amount covering the full text of one report | Test with materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Choose the value after labelling relevant and irrelevant samples | Takes effect only under semantic search or result reordering; compare against the labelled results |
| Question optimization | Enable first in multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the entity was altered |
| Result reordering | Enabled when a model is available and configured | Compare against the candidate order without reordering and observe the change |
Three common errors
- Treating a retrieval hit as authorization confirmation. Knowledge base search only returns reference content; it does not verify signature, approval, or effective status. Check: input a proposed judgment sentence containing the phrase "approved transfer," compare the retrieval result against the original transfer order, and observe whether an approval field appears. When it is absent, record it as pending verification.
- Treating semantic similarity as entity filtering. Minimum relevance is a filtering threshold and does not exclude unrelated entities or unauthorized records. Check: input a proposed judgment sentence concerning a different store, compare the entity of the recalled content against the target store, and observe whether material from another entity is mixed in. When it is, record it and add an entity check.
- Comparing values from different fields directly. Sales amount and transferred-in units belong to different fields and are not mutually exclusive values of the same field. Check: input the two original texts from the hypothetical sample above, compare the field names and the basis of the values, and observe whether entity, period, and field all match. When the fields differ, record the judgment as unsupported.
How to verify the setup
- Input the two original texts and the proposed judgment sentence from the hypothetical sample, compare entity, period, and field, and observe whether the judgment result is unsupported with the evidence gap retained. If so, the configuration is usable.
- Input one store material explicitly labelled relevant and one explicitly labelled irrelevant, compare the recall set at the chosen minimum relevance value, and observe whether the irrelevant material is filtered out. At the same time confirm that this threshold does not perform permission isolation. If so, the filtering setting is usable.
- Input an empty-reference case, compare whether the node output array length is 0, and observe whether the downstream path still runs. If so, the empty-result handling design is usable. All of the above are paper checks and do not claim that testing or deployment has taken place.