What the documents look like
Rebate calculation happens after sales close, and its center is policy interpretation and financial treatment. The working materials are rebate policies, sales attainment tables, reconciliation statements, and red-letter invoices. Source systems are ERP sales, dealer remittances, contract ledgers, and the finance system, with updates on a monthly or quarterly rhythm. Errors at this stage surface as rebate disputes, channel conflict, and incorrect revenue recognition.
For product data and selection questions, users ask in natural language whether a given model falls under a rebate tier, or which policy clause matches an attainment rate for a period. Model numbers, identifiers, and proper nouns in these materials need exact hits, while policy wording needs semantic matching. Those two needs sit side by side in the same knowledge base.
Constraints at the vector indexes step
Knowledge bases linked for retrieval must use the same vector model. Check the actual model and input type first; libraries carrying different vector model identifiers cannot be linked together. Record the identifiers, exclude the ones that do not match, and stop there.
One source data block can carry several indexes. Check the index entries and the original text content separately. Index content works best when it is complete and focused; retrieval coverage and index length need to be weighed against actual samples.
Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion time. Checking only the query side does not answer that question. Semantic recall and relevance scores do not enforce permission or entity isolation; retrieval is not authorization. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a switch that gets turned on. ReRank requires a model and must be enabled in settings. Image retrieval requires an actual image input and a model that supports it.
Hypothetical paper exercise sample: the source text reads, "Brand policy for the second quarter of 2025: for dealer A, model X1, when the remittance attainment rate from April through June is not lower than 90%, rebate is accrued at 2% of remittance amount; returned goods amounts are deducted from the attainment denominator." The draft sentence under review reads, "Dealer A has been confirmed to receive the second quarter rebate for model X1." The subject, object, and period in the draft match the material, but the material states only accrual conditions and does not state a confirmation outcome. The draft sentence is therefore unsupported and remains pending verification.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Vector model consistency | Linked libraries all carry the same vector model identifier | Documents require the same vector model; check identifiers library by library and exclude mismatches |
| Index content | Each block retains the full policy clause and model identifier | Balances semantic and exact hits; compare quoted source text using the same question set |
| Search mode | Hybrid retrieval | Model numbers need full-text hits; natural language questions need semantic recall |
| Minimum relevance | Hypothetical starting value 0.3 | Exercise starting point only; compare quoted source text against 0.2 and 0.4 |
| Reference limit | Hypothetical starting value 3000 tokens | Exercise starting point only; compare whether returned source text covers the clause |
| Result reranking | Enabled when a model is available | Documents state it can improve relevance; enable per settings and compare ordering |
Three common errors
- Linking libraries that use different vector models. Input: two libraries carrying different vector model identifiers. Compare whether the identifiers match. When they differ, exclude them and do not run cross-model retrieval.
- Treating a retrieval hit as rebate confirmation. Input: the draft sentence above. Compare whether the material states a confirmation outcome. When the material states only accrual conditions, the sentence is unsupported.
- Treating RRF as a switch that can be turned on. Input: a hybrid retrieval configuration. Compare whether an RRF toggle exists among the configuration items. The documents describe RRF as a fusion process, and no such toggle exists.
How to verify the setup
- Input the hypothetical source text and the draft sentence. Compare whether the material states "confirmed receipt of rebate." When only accrual conditions appear, record the result as unsupported and preserve the evidence gap.
- Input two libraries with different vector model identifiers. Compare whether the identifiers match. When they differ, record the exclusion result and make no claim about runtime behavior after linking.
- Input a query with no matching material. Compare whether the reference output is an array of length 0. When the output is an empty array, the downstream path still runs. Do not claim that an unrelated query must return empty.