What the documents look like
After-sales repair and reship work aims to restore a product to use. A case does not close through a direct refund decision. The working materials are repair applications, reship orders, inspection reports and accessory outbound notes. These records update per order and per day. Copy drafting draws checkable facts from them: fault descriptions, repair conclusions, accessory names and outbound status.
The materials sit in different systems. After-sales records sit in the after-sales system, order fields in the order system, inspection conclusions at repair sites, and accessory movements in the accessory warehouse. One order can draw fields from several systems. Before drafting, the drafter confirms which document and which version is being cited. A fault description in a repair application and a conclusion in an inspection report are separate fields, even when they concern the same order.
Constraints at the citation tracing step
The knowledge base search node accepts one or more knowledge bases that share the same vector model. It outputs cited content as an array, and the array length may be zero. A zero-length output still lets the downstream chain run. The retrieval result is handed to the answer step, where each drafted sentence is checked against the cited source text.
The minimum relevance value is a filter threshold. It does not act as a permission filter or an entity filter, and retrieval itself is not authorization. Mixed retrieval merges several recall paths through RRF. Reranking requires a model and runs when configured. Image retrieval requires image input and a model that supports the relevant capability. These optional paths do not all execute on every request.
Hypothetical paper exercise. On 12 March 2025 a brand after-sales site accepts a reship application for order A. The source text reads: "The inspection report states: water entered the main unit; replacement with a mainboard of the same model is recommended." and "The accessory outbound note states: mainboard outbound, quantity 1." A proposed copy sentence under review reads: "This repair is complete and the mainboard has been replaced." The source text supports only that the accessory left the warehouse. It does not state that replacement was completed, so this sentence is not supported. If the proposed sentence reads "the mainboard has been dispatched from the warehouse," it matches the outbound note on the same subject, the same period and the same field, and is supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | How to check |
|---|---|---|
| Linked knowledge bases | Select the knowledge base holding repair applications and inspection reports | Check the vector model identifier and exclude bases with a different model |
| Search mode | Mixed retrieval | Compare with semantic retrieval alone on queries containing identifiers |
| Minimum relevance | Hypothetical starting point 0.3 | Compare with 0.5 and observe whether unrelated passages decrease |
| Citation limit | Hypothetical starting point 2000 tokens | Compare with 1000 and observe whether citations are truncated |
| Reranking | Enabled | Compare with disabled and observe changes in candidate ordering |
| Reply format | Default | Change only when a downstream step needs a fixed structure |
The starting values are choices for this paper exercise, used for comparison with adjacent values. They are not official defaults. The thinking budget belongs to model thinking configuration and is unrelated to the citation token limit.
Three common errors
- Writing "repair is complete" from an accessory outbound note. The outbound note and repair completion are different fields. When no completion record exists, the status remains pending verification.
- Treating the minimum relevance value as a way to exclude unrelated business records. This value is a filter threshold. It does not recognize subjects or permissions, and unrelated records still require manual review.
- Treating an empty citation array as a workflow failure. The documented array length may be zero, and the chain still runs with an empty result. The downstream design for accepting empty output is what needs checking.
How to verify the setup
- Enter the inspection report text above together with the proposed sentence "This repair is complete and the mainboard has been replaced." Compare "replacement completed" against the source wording "replacement is recommended" and "outbound." When the source text does not state a completion status, the sentence is not supported. This is a paper check and has not been executed.
- Enter the accessory outbound note text together with the proposed sentence "the mainboard has been dispatched from the warehouse." Compare subject, date and field. When all three match, the sentence is supported. This is a paper check and has not been executed.
- Enter an empty citation array
[]and compare whether the downstream node is designed to accept an empty array. When the output is empty and the chain continues, the condition is met. This is a paper check and has not been executed.
These three checks are paper comparisons. They have not been run against a live system, and no deployment, approval or business action is claimed. Where the source text leaves a field unstated, the gap remains open and is recorded as pending verification rather than resolved.