What the documents look like
Regional dealer stock replenishment runs on daily rolling orders rather than seasonal centralized ordering, so the material set is continuous rather than event-driven. Four document families carry most of the review work: replenishment requests, inventory turnover tables, outbound delivery notes, and payment plans. They originate with different parties—dealer warehouses, brand channel managers, and sales back-office staff—and the same replenishment can appear in the request stage, the outbound stage, and the payment stage with field definitions that do not line up across those stages. The retrieval step locates the original text of these materials so a reviewer can compare it. Stockouts, channel overstock, and overdue receivables are consequences carried by business operations; retrieval does not prevent or cause them.
Because the same transaction is described at several stages, a review question frequently asks about status at one stage while the retrieved passage describes another. Retrieval returns candidate passages; it does not resolve which stage governs. That resolution stays with the reviewer reading the returned text.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the vector model identifier is checked first and non-matching bases are excluded before association. This is a configuration constraint on association, not a statement about what retrieval would return if the constraint were ignored.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval combines result sets through RRF, which is a fusion process applied to multiple recall paths rather than a switch that is turned on. Minimum relevance applies only under semantic retrieval or when result re-ranking is used; it is a filter threshold and does not perform permission or entity isolation. Retrieval is not authorization. The citation limit is measured in tokens and caps how much content a single search may cite. Question optimization performs coreference resolution and question expansion and adds an extra model call before retrieval. Citation content is returned as an array whose length may be zero, and the downstream chain still runs on an empty result.
Hypothetical paper exercise. Source text for comparison: "Dealer A in the East China region submitted a replenishment request in March 2025 for a quantity of 200 units; the status is pending review by the brand channel manager. No outbound delivery note has been generated for this request, and no payment plan has been entered." Draft sentence under assessment: "The replenishment request has been reviewed and approved, and shipment can be arranged." Assessment requires the same subject, the same period, and the same field. The source status field reads pending review; the draft sentence states approved. These are mutually exclusive values in the same field, so the draft sentence is unsupported. Pending review does not demonstrate an absence of operating authority, and it does not indicate that shipment has been executed.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the replenishment-related bases, after checking the vector model identifier | Bases on different vector models cannot be associated together; identifiers that do not match are excluded before association |
| Search mode | Run the sample once with semantic retrieval and once with full-text retrieval | Compare recalled content between modes against the test question; no single mode is presumed suitable for all material |
| Citation limit | Set a hypothetical starting point of 2000 tokens for this exercise | Test with material of differing lengths to see whether content is complete, then compare against the starting point and adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set a hypothetical starting point of 0.5 | This value applies only under semantic retrieval or result re-ranking; compare filtering behavior against the labeled results |
| Question optimization | Record the original question and the optimized question first | Check whether the subject was altered incorrectly, then decide whether to keep it for multi-turn follow-up scenarios |
| Result re-ranking | Use only when a re-ranking model is available and enabled by configuration | Re-ranking requires model support; when it is not enabled, no re-ranking score is produced |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input is an unrelated query; compare the recalled content against the authorized scope. The observable result is a change in relevance scores. The authorization boundary still requires downstream business system implementation, and retrieval results do not constitute authorization.
- Describing RRF as a switch that can be turned on. Input is a hybrid retrieval configuration; compare the semantic path and the full-text path. The observable result is a merged ordering. RRF is a fusion process and has no separate switch item.
- Reading a pending review status as approved. Input is the sample source text and the draft sentence; compare the status field values. The observable result is mutually exclusive values in the same field, and the draft sentence is unsupported. Approval completion cannot be inferred.
How to verify the setup
- Input the sample source text and the draft sentence; compare the status field values "pending review" and "approved." The observable result is that the two values are mutually exclusive in the same field, so the draft sentence is unsupported and the evidence gap remains open for verification.
- Input a query containing an explicit identifier and run it with full-text retrieval and with semantic retrieval; compare the recalled content from both paths. The observable result is that full-text retrieval matches the identifier while semantic retrieval is not guaranteed to match. Record the behavior of each mode; a first-result match is not guaranteed.
- Input a clearly paper-based empty result array []; compare whether the downstream chain continues to run. The observable result is that the output chain runs through. Use this to check the downstream design; do not claim that a fabricated rule search necessarily returns [].