What the documents look like
Creator commission settlement and commission deduction occur after a sale closes. The objects being reconciled are valid promotion orders, refund deduction details, and settlement bills. Commission setting records, settlement bills, refund deduction details, and invoices form the retrieval corpus. One settlement may appear in a bill and in a deduction detail under different language expressions for the same field. Retrieval needs to hit the same settlement object across languages. Semantic recall depends on vector distance. Identifier fields such as bill numbers and order numbers depend on full-text hits. Materials update weekly or monthly. The retrieval side only needs the indexed content to match the current settlement period's materials.
Constraints at the vector indexes step
Knowledge bases associated for retrieval must use the same vector model. Checking the actual model identifier and input type comes first. Cross-model retrieval is not executed to compare results. One original data block can link to multiple indexes. Index content and original text are checked separately. Index content should be complete and focused. Retrieval coverage and index length need to be weighed against samples. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion time, not only on the query side. Semantic recall and minimum relevance act on similarity only. They do not carry permission or entity isolation. Retrieval results do not constitute authorization. Hybrid retrieval merges multiple result paths through RRF. ReRank requires a configured model and must be enabled by settings. The reference limit is counted in tokens. Empty references output as an array and the length can be 0.
Configuration and implementation checks
The table below lists documented settings or implementation check items around this page's samples. Numeric values are hypothetical starting points for this trial, used for comparison against control values. They are not official defaults.
| Configuration or check item | Hypothetical starting point for this trial | Reason and comparison object |
|---|---|---|
| Vector model identifier | Assume both libraries register the same model identifier | Compare character by character against the other library's registered identifier; a mismatch excludes joint association |
| Index content | Each block retains settlement object, period, and field name | Compare against source text to confirm the index has not dropped fields needed for judgment |
| Reference limit | Hypothetical starting point 2000 tokens | Compare cited source text against the same question group at 1000 tokens |
| Minimum relevance | Hypothetical starting point 0.3 | Compare filtered results at 0.5 to observe missed recall and noise changes |
| Search mode | Semantic retrieval and full-text retrieval used together | Compare hit source text against semantic retrieval alone |
| Result rerank | Assume a rerank model is configured and enabled | Compare candidate order and relevance scores against the unreranked run |
Three common errors
- Associating knowledge bases that use different vector models. The input is two library registration identifiers. Compare whether they match; joint association is allowed only when they match. When they do not match, record the identifier difference. Do not infer runtime results.
- Treating semantic recall or a relevance score as a permission or entity filter. The input is the cited source text returned by one retrieval. Compare its subject against the requesting party's authorization scope. A citation beyond that scope is a retrieval hit and must be handled by downstream business verification.
- Writing RRF fusion as a toggle switch, or treating rerank as a mandatory step on every request. The input is the node configuration and whether this request contains an image and whether rerank is enabled. Compare the actual execution path against the optional paths listed in the documentation. When not enabled, that path produces no result.
How to verify the setup
Hypothetical drill sample: a settlement specialist needs to reconcile one creator's March 2024 commission. Source text provided for reconciliation: "Settlement bill SB-2024-0312 states that creator A had 120 valid promotion orders in March 2024, with a commission base of 8000 yuan; refund deduction detail TK-2024-0315 states that 15 refund orders occurred in the same period, with a commission deduction of 900 yuan." The draft sentence to be judged: "This creator's March 2024 settlement commission is 8000 yuan." The judgment input is the two source passages above and the draft sentence. Compare period, subject, and field values. The draft sentence does not deduct the refund commission and is unsupported. A correct statement must cite both source passages and state the deduction relationship.
- Check one: the input is the two libraries' vector model registration identifiers and the sample source text above. Compare whether the identifiers match. When they match, associate jointly and retrieve "SB-2024-0312 deduction". Observe whether the cited source text hits both the bill and the deduction detail. Hitting only one is a missed recall; record the gap.
- Check two: the input is the cited source text for the same question at minimum relevance 0.3 and 0.5. Compare the bill number and deduction amount fields returned in both runs. Observe whether both source passages are still retained at 0.5. If the deduction detail is filtered out, record that field as missing.
- Check three: the input is one retrieval request with no image uploaded and rerank not enabled. Compare the actual returned path against the optional paths listed in the documentation. Observe whether only text recall runs. When the reference array is empty, use the paper input [] to check whether the downstream node still proceeds. Do not infer that the retrieval rule necessarily returns an empty result.