What the documents look like
Price and display auditing takes place during day-to-day operations, where the focus is sustained execution of standards. For procurement and supplier material, the documents entering the knowledge base are typically store visit photographs, price tag lists, display standards, rectification notices, supplier quotations and supply confirmation records. These materials update weekly or monthly, and the working roles are the chain operations department, regional supervisors and store managers.
The retrieval step faces material in which subject, object, status and validity scope are mixed within a single document. A store visit photograph may cover only one store on one day. A quotation may be valid only for a defined period. A confirmation receipt may record a received document and a pending review rather than an approval. Retrieval results must let the user see these boundaries before any downstream comparison is made. Retrieval itself does not grant authorization, and a similarity score does not act as a permission or entity filter.
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 built on different vector models cannot be associated together. When identifiers differ, the correct action is to check the identifier and exclude the non-matching base, not to run a cross-model search in order to compare missed recalls.
Search modes are semantic search, full-text search and hybrid search. Hybrid search merges multiple result paths through RRF, and RRF is a fusion process rather than a switch. Reranking requires a model and must be enabled in settings. Image retrieval requires image input and the corresponding model capability. The optional paths shown in the overall document flow diagram do not mean every request executes all of them.
The reference limit is measured in tokens. The minimum relevance value applies only during semantic search or when result reranking is used. It is a filtering threshold, not a final ordering rule.
Hypothetical paper exercise: on 10 March 2026 a regional supervisor of a chain brand submits a draft sentence for review: "Supplier A will apply the quotation price to East China stores from 1 March 2026 to 31 March 2026, and the quotation has been confirmed by the procurement department." The source material provided for comparison reads: "Supplier A quotation (number HYP-0310) states a validity period from 1 March 2026 to 31 March 2026, applicable to East China directly operated stores; the procurement department receipt dated 5 March 2026 states 'quotation received, pending review by the price committee'." In this sample the subject is Supplier A, the object is East China directly operated stores, the status is pending review, and the validity scope is 1 March to 31 March.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Associate one knowledge base containing supplier quotations and receipts | Check the vector model identifier of that base first; include it only when it matches the retrieval side, exclude it when it does not |
| Search mode | Hybrid search | The sample contains both the identifier HYP-0310 and natural language wording; compare recall differences between full-text and semantic paths |
| Reference limit | Set a token value that covers the full quotation text | Use complete inclusion of the quotation and the receipt as the comparison baseline; content is truncated when the value is insufficient |
| Minimum relevance | Start with a middle value, then compare against labelled samples | Applies only during semantic search or reranking; observe the filtering boundary using relevant and irrelevant samples |
| Question optimization | Enable for follow-up questions containing references | Compare the original question with the optimized question, and check whether subject and period were altered incorrectly |
| Reranking | Enable per settings when a model is available | Observe ordering using the rerank score and the vector score separately, and confirm the candidate set is sufficient |
Three common errors
- Treating "pending review by the price committee" as confirmed. The receipt states only "quotation received, pending review", while the draft sentence writes "confirmed by the procurement department". This upgrades a pending status into a confirmed one. Check: input the receipt text and the draft sentence, compare the status words "pending review" and "confirmed", and observe whether they point to the same fact.
- Treating the quotation validity period as a supply period. The sample states only a quotation validity period and does not state a supply period. Check: input the quotation text and the draft sentence, compare whether "validity period" and "supply period" are the same field, and observe whether the draft sentence introduces a period conclusion not stated in the source.
- Associating knowledge bases across different vector models. Check: input the vector model identifier of the base to be associated, compare it item by item with the identifier on the retrieval side, and observe whether any base with a mismatched identifier is still included in the association.
How to verify the setup
- For the sample draft sentence, input the quotation and receipt text, and compare "pending review" with "confirmed". If the retrieved receipt text clearly states "pending review", the draft sentence is not supported and should be recorded as pending verification.
- For the sample period, input the draft sentence and the quotation text, and compare whether "1 March 2026 to 31 March 2026" appears both as the quotation validity period and in the draft sentence. If the quotation states only a validity period, the supply period remains a gap.
- For empty-result handling, use a clearly paper-based empty array input to check the downstream design, and observe whether the downstream path proceeds as a reference chain whose length may be zero. Do not claim that a fictional rule search necessarily returns an empty array.
All three items use the hypothetical exercise sample as input. The expected observation is a record of unsupported content or a gap. This does not indicate that testing, approval or acceptance has been completed.