What the documents look like
After-sales repairs and reships aim to restore a product to working condition. The case does not close through a direct refund, so the paper trail follows repair applications, reship orders, inspection reports and accessory outbound documents. These materials come from the after-sales system, the order system, repair sites and the accessory warehouse WMS. In cross-border work, the same case may involve documents written in different languages, and a search result needs to point back to a specific document and version.
Updates occur per order and per day. A citation should correspond to the material state of that order on that day. Repair or reship work also has a stage structure: a reship order can be pending review, a signature field can be blank, and an inspection report can carry a recommendation. Those are separate facts. A search node returning a document does not mean the document was approved, signed or executed.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Its output is an array of citations, and the length can be zero. When no content is found, that output path still runs. The minimum relevance value takes effect only under semantic retrieval or result reranking, and it acts as a filter threshold. It does not perform permission or entity isolation. Hybrid retrieval combines multiple result paths through RRF. Reranking requires a model and is enabled according to settings. Image retrieval requires image input and a model with the corresponding capability. These paths do not all execute on every request.
Hypothetical paper exercise sample: an accessory warehouse at a brand after-sales operation receives a reship order. The order number is assumed to be BX-2024-0001, the date scope is assumed to be 1 March 2024 to 31 March 2024, the object is a motor accessory of a given model, the status field reads "pending review", and the signature field is blank. A separate inspection report, number JC-2024-0001, has a conclusion field reading "reship recommended". A draft sentence for review reads: "The reship order has passed review and can be scheduled for outbound." When checking, the reship order status field and the draft phrase "passed review" belong to the same field, and their values are mutually exclusive, which forms a contradiction. The blank signature field shows that authorization is not complete. Signature, review and outbound are three distinct facts. The retrieval scope must cover both materials; when a material is missing, the evidence gap is recorded.
Configuration and implementation checks
| Check item | Assumed starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Two bases with the same vector model only | Compare against the identifier of a base with a different vector model; exclude non-matching ones |
| Search mode | Hybrid retrieval | Compare with pure semantic retrieval for numbered-item matching |
| Minimum relevance | Assumed starting point 0.5 | Compare with 0.3 and observe the change in the number of citations after filtering |
| Citation limit | Assumed starting point 2000 tokens | Compare with 1000 tokens and observe truncation |
| Result reranking | Enabled | Compare with disabled for ordering stability |
| Citation output handoff | Citation array passed to the answer step | Compare with a direct answer for whether each item can be traced back |
These values are assumed starting points for this exercise and are not official defaults.
The configuration table keeps the checks tied to the sample materials. The two knowledge bases must share a vector model before they can be associated together. The citation limit uses the documented token unit. The minimum relevance value is a filter threshold and does not decide which business records a user may see. Reranking and image retrieval are optional paths that depend on model capability and settings, and they are not guaranteed to run on every request.
Three common errors
- Treating retrieval or a form as authorization execution. The input is a reship order with status "pending review", compared against the draft sentence "passed review". Scheduling outbound directly is an error. The correct handling records that authorization is not complete.
- Treating minimum relevance as a permission or entity filter. The input is a cross-entity query, compared against the returned citations. Concluding that the similarity setting has excluded unrelated records is an error. Retrieval does not perform authorization.
- Associating knowledge bases with different vector models. The input is the identifiers of two bases with different vector models, compared against the association condition. Forcing the association is an error. The identifiers should be checked first, non-matching ones excluded, and no runtime result inferred.
How to verify the setup
- Input the reship order BX-2024-0001 and the inspection report JC-2024-0001 from the hypothetical sample. Compare the draft sentence "The reship order has passed review and can be scheduled for outbound" against the source materials. The check is satisfied when the retrieved citations can be traced to the status field "pending review" and the blank signature field.
- Input the same two materials. Compare the answer citations against the source text. The traceability condition is satisfied when each citation can be located to its corresponding document and field. When a material is not retrieved, record the gap.
- Input an explicit paper-level empty array []. Compare the downstream handoff design. The empty-result handoff condition is satisfied when the downstream step continues without error. Do not claim that a fictional rule will always return [].
These checks describe expected observations. They do not claim that a test, deployment, approval or outbound action has taken place.