What the documents look like
Regional dealer stock replenishment runs as a rolling daily order flow. It differs from seasonal centralized ordering events. The main data comes from dealer ERP systems, brand ordering systems, warehouse inventory, and sales collections. Typical actors are dealer warehouses, brand channel managers, and sales back-office staff. The everyday document set consists of replenishment requests, inventory turnover sheets, outbound delivery notes, and collection plans, with weekly or monthly update frequency. These materials are often transferred together during an operations handover, and the retrieval step must separate the material itself from its authorization state.
Two or three facts shape this page. Replenishment requests carry a review status that can remain pending while the request is already stored and searchable. Outbound delivery notes carry a separate receipt-confirmation status, so retrieval of a delivery note does not mean the goods were signed for. Collection plans describe intended payments, not received funds. Each of these states lives in its own field, and retrieval returns text, not a verified business state.
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 checking, confirm the model identifier first and exclude the non-matching bases. Do not run a cross-model retrieval to compare missed recalls.
Semantic search, full-text search, and hybrid search are compared separately against test questions. No single mode is assumed suitable for all materials. The citation limit is measured in tokens, so test with actual materials of differing length to see whether content returns complete. The minimum relevance value takes effect only under semantic search or when result reranking is used. Label relevant and irrelevant samples before adjusting the threshold. Question optimization performs coreference resolution and question expansion, so compare the original question with the optimized question and check whether the subject was altered. Retrieval and similarity scores do not perform permission or entity isolation, and a retrieval result does not constitute authorization.
Hypothetical paper exercise sample: a brand channel manager hands over a replenishment request, number B-0714, whose source text reads "Applicant: East China Region 3 dealer; replenishment item: Type A filter; requested quantity: 200 units; request date: 14 July 2025; current status is pending review by brand channel manager." Attached is an outbound delivery note, number C-0715, whose source text reads "Outbound item: Type A filter; outbound quantity: 200 units; outbound date: 15 July 2025; receipt status: not signed." The draft sentence to be judged reads "B-0714 has been approved, and C-0715 has been signed for." This draft sentence conflicts with the field states in the two source texts.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases using the same vector model | Sample materials sit in a same-model base; check identifiers and exclude cross-model bases first |
| Search mode | Compare semantic, full-text, and hybrid search per test question | Numbers B-0714 and C-0715 suit full-text or hybrid checking |
| Citation limit | Unit is tokens; test with actual materials of differing length | Sample holds two passages; confirm both return complete |
| Minimum relevance | Takes effect only under semantic search or result reranking | Label relevant and irrelevant sample fragments before adjusting |
| Question optimization | Performs coreference resolution and question expansion | Compare original and optimized question; check subject unchanged |
| Result reranking | Requires a rerank model and must be enabled per settings | Enable only when the sample has many candidates; not run by default every time |
Hypothetical trial starting value for this page: the citation limit is set to 512 tokens. The reason is that the two source passages together run roughly 200 characters, so 512 covers both with margin. Compare against 256 and observe whether the second passage is truncated. This value is a hypothetical starting point for this exercise, not an official default.
Three common errors
- Treating a retrieval hit as completed authorization. In the sample, B-0714 is "pending review by brand channel manager." Retrieving that request does not mean the review passed, and it does not mean C-0715 was signed for. The error is substituting the retrieval result for a field-state check.
- Associating knowledge bases built on different vector models. If the sample materials were spread across bases with different vector models, check the identifiers first and exclude the non-matching ones, then record what was actually observed. Do not invent missed recalls or redundant results from a joint association.
- Reading a pending review as missing operational permission. B-0714 in pending-review state only shows the review action is incomplete. It does not show the applicant lacks replenishment permission. Authorization state and process state are checked separately.
How to verify the setup
- Input the sample source text "B-0714 current status is pending review by brand channel manager" together with the draft sentence "B-0714 has been approved." Compare the value of the same field, status: the source says pending review, the draft says approved. The observed result is that the two values are mutually exclusive, the draft sentence is not supported by the source, and it should be recorded as pending verification. If retrieval returns an empty array, the downstream path follows the documented empty-reference design; do not claim a rule-based retrieval must return an empty array.
- Input the sample source text "C-0715 receipt status: not signed" together with the draft sentence "C-0715 has been signed for." Compare the value of the same field, receipt status: the source says not signed, the draft says completed. The observed result is that the two values are mutually exclusive and the draft sentence is unsupported. Retrieving the delivery note does not mean receipt is complete; receipt and outbound are different facts.
- Input both sample source passages and check whether the content returned under a 512-token citation limit contains both the B-0714 passage and the C-0715 passage. Compare the returned content against the source passages and observe whether the second passage is complete. If only the first passage returns, record it as insufficient citation limit or a chunking issue, adjust the limit, and compare again. Do not guarantee a first-hit or a full recall.