What the documents look like
Dealer rebate calculation happens after sales close. The work centres on checking policy wording, confirming achieved volume, deducting returns and completing the financial offset. In store service and internal training, staff explain the basis of a rebate figure and state which document each conclusion comes from.
Four document types form the material base for citation tracing. The rebate policy sets the calculation rule and the applicable period. The volume achievement table records actual sales quantities. The reconciliation statement confirms the amount both sides accept. The red-letter invoice completes the billing offset. Materials are updated monthly or quarterly, and the version used in store training needs to match the calculation period.
Two features of this category shape the tracing task. Rebate calculation sits after the sale, so the evidence chain runs from policy condition to achieved volume to settled amount. The documents are period-bound, so a citation that does not carry its period cannot be matched to the training question.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model, for vector search. The search node outputs cited content as an array, and the length can be zero. Even when no content is found, this output path still runs through.
Citation tracing requires checking each answer statement against the cited source text, and confirming the cited content comes from a valid version inside the selected knowledge bases.
Hypothetical drill sample: a brand finance department holds a March 2025 rebate policy stating that a dealer reaching quarterly volume of 500 units is rebated at 200 yuan per unit. A March 2025 volume achievement table records that a dealer actually sold 480 units. A draft sentence awaiting review reads: "This dealer reached the quarterly volume target in March 2025 and can receive a rebate at 200 yuan per unit." The review question is whether this draft sentence matches the cited rebate policy and volume achievement table. The policy condition is 500 units; the achievement table records 480 units, which does not meet that condition. The conclusion in the draft sentence does not match the two source documents.
Retrieval output is not an approval. The search node returns a citation array; the retrieval action itself does not constitute rebate approval, billing offset or payment authorisation. A relevance score is a filtering threshold, not a permission filter, and it does not exclude unauthorised records. The same-vector-model requirement applies to associating knowledge bases inside the search node; it does not extend to any external system.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select the rebate policy base and the volume achievement base | Both bases must use the same vector model to be associated together; compare the vector model identifiers for consistency |
| Search mode | Hybrid search | The rebate policy contains codes and amounts, suited to full-text search; store questions vary in wording, suited to semantic search |
| Citation limit | Hypothetical starting point of 3000 tokens | Intended to hold policy clauses and achievement table fragments; compare against the length of a single rebate policy clause and adjust |
| Minimum relevance | Hypothetical starting point of 0.4 | Intended to filter low-relevance fragments; compare against whether cited content covers the policy condition sentence and adjust |
| Question optimisation | Enabled | Store training has many follow-up questions, used to complete references; compare against retrieval results with it disabled |
| Result reranking | ReRank enabled | Used for secondary ordering when candidates are numerous; compare against ordering with it disabled |
These values are hypothetical trial starting points for this drill, not official defaults. The configuration items follow the settings described in the knowledge base search node and the knowledge base search scheme documents.
Three common errors
- Treating retrieval output as an authorisation or approval result. The search node outputs a citation array; the retrieval action does not constitute rebate approval, billing offset or payment authorisation. Error symptom: training material states "the system has confirmed the rebate amount" when only retrieval citations exist. Check: input is the retrieval citation array, compare whether an approval record exists alongside the cited content, and observe that when the citation array exists but no approval record exists, the wording must not state that the amount is confirmed.
- Judging an unmet volume record as met. Error symptom: the draft sentence states "reached the quarterly volume target" while the volume achievement table records 480 units and the rebate policy condition is 500 units. Check: input is the rebate policy source text and the volume achievement table source text, compare the two values 480 and 500 and the condition wording, and observe that when the value does not meet the condition, the draft sentence is judged unsupported.
- Associating knowledge bases that use different vector models. Error symptom: a policy base and an achievement table base using different vector models are both selected into one search node. Check: input is the vector model identifiers of the two knowledge bases, compare whether the identifiers are consistent, and observe that when the identifiers differ, exclude the non-conforming base and do not run cross-model retrieval.
How to verify the setup
- Check cited content against source text item by item. Input is the citation array output by the search node together with the rebate policy and volume achievement table source text. Compare the condition value in the cited fragment against the conclusion in the draft sentence. Observe that when the cited fragment contains the 500-unit condition while the draft sentence states the target is reached, the result is judged inconsistent. Expected outcome: cited content can be traced to a specific document and clause.
- Check how an empty citation output is carried downstream. Input is a paper-set empty citation array []. Compare the downstream answer step's handling design when it receives an empty array. Observe whether the downstream design distinguishes an empty citation from a retrieval error. Expected outcome: the path still runs through on an empty array, and it is not judged a retrieval failure.
- Check material version against period. Input is the March 2025 rebate policy and the March 2025 volume achievement table. Compare the period markers of the two materials against the period covered by the store training question. Observe that when the periods match, the citation is valid, and when the periods differ, record it as pending verification. Expected outcome: citation source corresponds to the calculation period.
These verifications are paper review designs and were not actually executed. Retrieval results are affected by data quality, the vector model and the search terms, and first-hit or full recall is not guaranteed.