What the documents look like
Regional dealer stock replenishment runs on daily rolling orders rather than seasonal central ordering. The working documents are replenishment requests, inventory turnover tables, outbound notes and payment plans. These records come from dealer ERP systems, brand ordering systems, warehouse stock and sales receipts, and they update weekly or monthly. In cross-border work, the same batch of material may exist in both Chinese and English versions, so a retrieval result has to show which version and which point in time was matched. Document timeliness decides whether a citation can be used: when last month's inventory turnover table and this week's replenishment request are both recalled, the answering step has to check the date range of each cited passage one by one. A replenishment request is a submitted document, not an approval record, and a retrieval hit on that request does not carry any authorization state with it.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. That zero-length output means the output path still runs even when nothing was found. The same-vector-model requirement only limits which knowledge bases may be associated together; it does not extend to business registration sites or other external systems. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result reranking is used, and it does not act as an entity or permission filter. RRF is a fusion step that combines multiple recall paths, not a switch to be turned on. Optional paths in the overall retrieval flow, such as image retrieval or reranking, run only when the corresponding input, model capability and settings are present. Citation content has to be passed to the answering step, where each answer sentence is checked against the cited source text.
Hypothetical paper exercise sample: a regional dealer submits a replenishment request on 10 March 2026. The request text reads: "This company requests replenishment of 200 units of model A stock, expected to arrive at the warehouse before 5 April, with the 8 March outbound note and the February payment plan attached." A draft sentence reads: "The dealer was approved on 10 March to replenish 200 units." The evidence needed for judgment is the request text itself, which records only submission and states no approval status and no approved quantity. The phrase "was approved" has no corresponding field in the source text, so the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only replenishment material bases that share one vector model | Check each base identifier against the vector model identifier, and exclude any that do not match |
| Search mode | Hybrid retrieval | Compare the number of citations against semantic retrieval alone |
| Minimum relevance | Hypothetical starting value 0.5 | Compare the change in filtered citation count against 0.3 and 0.7 |
| Citation limit | Hypothetical starting value 2000 tokens | Compare against 1000 and 4000 to see whether cited content covers the source text |
| Result reranking | Enable when a rerank model is available | Compare citation ordering before and after enabling |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare how well cited source text matches the question before and after enabling |
These values are hypothetical starting points for this exercise, used to compare citation behavior at different settings. They are not official default values.
Three common errors
- Treating a retrieval hit as completed authorization. The replenishment request text records only the act of submission. Finding that request does not mean approval has passed. Citation tracing has to keep submission, review and approval as three separate states, and an unsigned or pending item is not a completed action.
- Treating minimum relevance as an entity or permission filter. The threshold only filters low-relevance results. It cannot exclude records that belong to a different dealer. Citation checking still has to return to the subject field in the source text. A relevance score is not a category, entity or permission filter.
- Treating an empty citation array as a workflow failure. The search node output length may be zero, and that empty array has to be handled by downstream design. It cannot be read as proof that the retrieval rule always returns empty or that the system has failed. Without a documented error code or an actual observation, the failure cause remains unknown.
How to verify the setup
- Input the replenishment request text from the hypothetical sample together with the draft sentence "The dealer was approved on 10 March to replenish 200 units." Compare the approval status in the draft sentence against the fields stated in the source text. When the source text states no approval result, the judgment is unsupported. It is not a contradiction, because contradiction requires the same subject, the same period and the same field with mutually exclusive values.
- Input the Chinese and English versions of the same batch of replenishment material. Compare the version identifier and date range matched in the citation output. The cited content has to correspond to a specific version. When a version identifier is missing, record the evidence gap rather than filling it in.
- Input an explicitly constructed empty paper result [] and compare whether the downstream answering step handles it as zero citations. If the downstream step treats the empty array as a normal no-citation state, the condition is met. If it treats the empty array as an error and stops, record the observed behavior. Do not assert that the system necessarily reports an error. A pending request is not approval, an offered option is not execution, and retrieval is not authorization. Business verification, approval, database writes and external synchronization all need their own stated downstream implementation.