What the documents look like
Repair and reship work aims to restore a product to use. It does not close a case through a direct refund, which is the pattern in refund-only disputes. For store service and internal training, the working materials are the repair application, the reship order, the inspection report, and the accessory outbound order. These come from the after-sales system, the order system, the repair site, and the accessory warehouse WMS. At handover between store staff and trainers, citation tracing must answer a narrow question: which material, in which version, at which status, supports a given statement or request. Materials update per order and per day, so store staff check the current version against the material they cite.
A citation is a pointer to a source. It is not an authorization record, an approval, or proof that a business step has been executed. The knowledge base search node outputs a reference array whose length may be zero, and the output path still completes when nothing is found. That empty array cannot be read as a business failure or as permission to proceed.
Constraints at the citation tracing step
Retrieval and citation do not carry authorization authority. The knowledge base search node can associate one or more knowledge bases that share the same vector model. Knowledge bases with different vector models cannot be associated together. Hybrid retrieval merges multiple result paths through RRF, and result reordering requires a reordering model enabled by settings. Image retrieval requires image input and matching model capability. These are optional paths; their presence in the documented flow does not mean every request executes all of them.
In the AI dialogue node, the reply format can be set to jsonschema when a downstream step must read fixed fields. Thinking configuration follows the documented effort enumeration and aligns with reasoningeffort. A thinking budget and a reference token cap are different quantities and must not be substituted for each other. The reference cap is expressed in tokens, not in a guessed item count.
Hypothetical drill sample. The source material reads: "Repair site A received the repair application for order B on 10 March 2025. Inspection report number C states that the mainboard was affected by moisture and recommends reshipping the same model accessory. Accessory warehouse WMS shows that outbound order D was generated on 12 March 2025 with status pending review." The proposed determination sentence is: "The reship is complete, so the store can confirm with the customer." Determination points: outbound order D shows pending review and does not state completion. The inspection report only recommends a reship and does not state that a reship was executed. The proposed sentence writes a pending-review status as completion, which the source does not support.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select knowledge bases holding repair applications, reship orders, inspection reports, and accessory outbound orders that share one vector model | Matches the material version; excludes knowledge bases on a different vector model |
| Search mode | Hybrid retrieval | Numbered lookups and natural-language lookups coexist; compare against semantic retrieval alone |
| Reference cap | Hypothetical starting value set to cover one inspection report and one outbound order in tokens | Compare against a lower value covering only one material and observe reference completeness |
| Minimum relevance | Hypothetical starting value at the middle value described in the documentation | Compare against a high value and observe whether numbered materials are missed |
| Result reordering | Enabled when a reordering model is configured | Compare against leaving it off and observe changes in candidate order |
| Reply format | json_schema when a fixed structure is required | Compare against the default format and judge whether downstream steps can read fields |
These values are hypothetical starting points for this drill and are not official defaults.
Three common errors
- Reading pending review on outbound order D as completion and writing that the reship was executed. The input is the status field of outbound order D. The comparison is between "complete" and "pending review." The observable result is that the proposed sentence conflicts with the source status.
- Reading the recommendation in inspection report C as an approved reship. The input is the inspection report text. The comparison is between "recommends" and "approved." The observable result is that the proposed sentence lacks an approval basis.
- Treating a retrieved citation as an authorization. The input is the output array of the knowledge base search node. The comparison is between the reference content and any authorization record. The observable result is that a reference exists while authorization material is absent.
How to verify the setup
- Enter the drill source text and the proposed determination sentence. Compare the status field of outbound order D against the wording "complete." If the proposed sentence still states completion, the check fails.
- Enter the text of inspection report C. Compare "recommends reshipping" against "approved." If the proposed sentence does not mark the point as pending verification, the check fails.
- Enter the case where the knowledge base search node outputs an empty array, written as []. Compare whether the downstream design handles a length of zero. If the downstream step treats the empty array as an error and stops, the check fails.