What the documents look like
Dealer rebate calculation happens after sales, and its core is policy interpretation and financial treatment. The working documents are rebate policies, sales achievement tables, reconciliation statements, and red-letter invoices, refreshed monthly or quarterly. In cross-border settings the same set of materials may appear in different languages, so brand finance, channel sales, and dealer finance need to locate the same policy clause or the same reconciliation basis across languages. Semantic retrieval supports cross-language understanding, and matching a Chinese question to an English knowledge point is a documented capability. Field definitions in rebate materials must be aligned item by item for entity, period, and field. Retrieval only brings candidate source text forward; it does not determine authorization or effective status.
The knowledge base is composed of libraries, collections, and data. The smallest search unit is the library; collections classify data and do not affect search results. One source data block can be linked to multiple indexes, and the index and the source content are checked separately. Knowledge bases linked for retrieval must use the same vector model, so the actual model and input type are verified first, and libraries with different vector models cannot be linked together. Index content is built from complete and focused information, and retrieval coverage against index length is weighed against the sample. Image retrieval depends on whether an image description or image vector index was generated at ingestion time, and cannot be judged from the query side alone. Hybrid retrieval merges semantic and full-text results through RRF, and RRF is a fusion process. The minimum relevance value takes effect only in semantic retrieval or result reranking, and it is a filter threshold. The citation limit is controlled by tokens. Semantic recall and relevance do not carry permission or entity isolation functions, and retrieval results do not constitute authorization.
Constraints at the vector indexes step
Linking knowledge bases with different vector models is outside the documented condition. The check is on the model identifier and input type, and non-matching candidates are excluded and recorded; a cross-model retrieval is not run to compare missed recall. The minimum relevance value filters low-relevance results and does not exclude records belonging to other entities, so permission requires a separate downstream implementation. A retrieval hit is not evidence that a rebate was approved or invoiced; the source text may not state approval or invoicing status, and that fact remains to be verified.
The citation limit uses the documented token unit rather than a guessed item count. The minimum relevance value retains its documented condition of applying only in semantic retrieval or with result reranking. An empty-reference output is supported: the output is an array whose length can be zero, and the downstream path still completes. Image retrieval requires an actual image input and the corresponding model capability, and reranking requires a model that is enabled in the configuration. The optional paths in the overall retrieval flow are not executed on every request.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Vector model of linked knowledge bases | Hypothetical starting value: one shared model identifier | Documents require the same vector model; verify the identifier and exclude non-matching libraries, comparing candidate libraries one by one |
| Search mode | Hypothetical starting value: hybrid retrieval | Rebate policies contain both clause semantics and identifiers; hybrid retrieval uses semantic and full-text recall together |
| Index content | Hypothetical starting value: segmented by policy clause | Complete and focused indexes raise precision; compare recall coverage against a single long-text index |
| Citation limit | Hypothetical starting value: one token value | Chunk lengths differ widely; compare result stability against a top-k approach |
| Minimum relevance | Hypothetical starting value: enabled only for semantic retrieval or reranking | Documents state the value takes effect only in these two cases; compare candidate counts with the value disabled |
| Image index | Hypothetical starting value: image description generated at ingestion | Image retrieval depends on the ingestion-time index; compare hit situations against uploading an image only on the query side |
Three common errors
- Linking libraries that use different vector models. Input: two libraries whose model identifier fields differ. Comparison: the model identifier fields. Observable result: the identifiers do not match, so the non-matching library is excluded and recorded, and no cross-model retrieval is run to compare missed recall.
- Treating the minimum relevance value as a permission or entity filter. Input: a low-relevance rebate source text that belongs to another entity. Comparison: the documented conditions under which the value takes effect. Observable result: the value filters low-relevance results only and does not exclude records of other entities; permission requires a separate system implementation.
- Treating a retrieval hit as an approved or invoiced rebate. Input: a retrieved rebate policy source text. Comparison: its status field. Observable result: the source text does not state an approval or invoicing status, and that fact remains to be verified.
How to verify the setup
Hypothetical exercise sample: the source material reads, "East China region dealer rebates for the third quarter of 2025 are calculated in tiers by quarterly sales achievement rate; the achievement rate is based on ERP sales data, and return amounts are deducted before calculation." The draft sentence to be assessed is, "This dealer's third-quarter rebate has been calculated and approved under the above policy." The entity is a dealer in the East China region, the object is the third-quarter rebate, the status is draft, and the effective scope is the third quarter of 2025. The source text does not state a calculation completion or approval status.
- Verification one: input the source text above and the draft sentence; compare whether the draft sentence exceeds the scope stated in the source text; the observable result is that the source text describes only the calculation basis and does not state calculation completion or approval, so the draft sentence is not supported.
- Verification two: input one rebate policy source text and one reconciliation statement source text; compare whether entity, period, and field match; the observable result is that only the same entity, the same period, the same field, and mutually exclusive values constitute a contradiction, while evidence from a related stage alone remains to be verified.
- Verification three: input an empty reference array; compare the downstream node design; the observable result is that the reference length can be zero and the output path still completes, without claiming that a retrieval of a fictional rule necessarily returns an empty array.
After a model or index adjustment, compare cited source text using the same set of questions, and do not promise an effect from a vector score alone. At operational handover, the authorization status of materials, operations, and requests must be checked separately, and retrieval and forms do not constitute authorized execution.