What the documents look like
Concession counters in shopping malls take payment through the mall's central checkout, and the brand settles with the mall under a concession deduction arrangement. The same sale therefore appears in the mall POS, the brand ERP, the finance system, and the monthly reconciliation statement, each with its own purpose and level of detail. Product information and selection questions draw on four document types: concession deduction contracts, mall activity letters, sales receipts, and monthly reconciliation statements. The people handling this material are brand retail staff, mall operations staff, and concession counter managers. These documents update daily and monthly.
For retrieval, contract clauses, activity conditions, receipt line items, and reconciliation amounts sit in separate materials. A vector model and index must connect fragments that describe the same product or the same settlement item back to a checkable original passage. The smallest search unit is the library, not the collection; collections exist for classification and do not affect search results.
Constraints at the vector indexes step
Libraries associated for one search must share the same vector model. Before linking them, confirm the actual model identifier and its supported input types, and exclude any library that does not match. Libraries built on different vector models cannot be associated together. This is a check performed before association, not a comparison run across models.
In MongoDB's dataset.datas table, one original data record can correspond to multiple vector IDs held in an indexes field. PostgreSQL stores vectors in a vector field. During retrieval, vectors are recalled first, then the original content is fetched from MongoDB by vector ID; records mapping to the same original data are merged and the highest score is kept. Index content should be complete and focused, and the trade-off between retrieval coverage and index length depends on the sample at hand.
Image retrieval depends on whether an image description index or an image vector index was generated at ingestion time. Uploading an image on the query side does not by itself make an image retrievable. Hybrid search merges semantic and full-text results through RRF. Reranking requires a configured model and must be enabled under its settings. The minimum relevance score applies only in semantic search or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. Retrieval is not authorization. Forms collect values; business verification, approval, database writes, and external synchronization each need their own downstream implementation.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to the sample on this page. The numeric values are hypothetical starting points for this exercise, used to compare against the cited original text for the same question group. They are not official default values.
| Configuration item | Hypothetical starting point for this exercise | How to check and why |
|---|---|---|
| Associated libraries | Select only libraries with a matching vector model identifier | Check the model identifier and input types first, exclude non-matching libraries, and do not run cross-model retrieval comparisons |
| Search mode | Hybrid search | Contract clauses and activity letters contain numbers and conditions, so semantic and full-text paths are combined through RRF |
| Result reranking | A rerank model is configured and enabled under its settings | The rerank score is used for filtering; this path is not enabled when no model is configured |
| Reference cap | Hypothetical starting point of 2000 tokens for this exercise | The cap controls reference volume in tokens; compare it with the actual reference length for the same question group |
| Minimum relevance | Hypothetical starting point of 0.35 for this exercise | Applies only in semantic search or reranking; compare results before and after filtering for the same question group |
| Image index | Generated at ingestion according to model capability | Check whether an image description index or image vector index was built; an upload on the query side does not replace an ingestion index |
Hypothetical exercise sample. The original material for checking reads: "Concession deduction contract, Clause 3: the counter's sales for the month are settled at a 22% deduction. Mall activity letter No. 2024-06: a full-reduction activity runs from 15 June to 30 June, and the deduction follows the original ratio. Monthly reconciliation statement: the June deduction amount is based on the mall POS summary." The draft sentence or handover request to be judged is: "Please confirm the deduction ratio during the 15 June activity period and adjust the reconciliation statement accordingly." The original text contains the subject, object, time range, and settlement basis needed for the check.
Three common errors
- Associating knowledge libraries that use different vector models. The input is two libraries with non-matching model identifiers. Compare the model identifiers and input types, and observe whether the non-matching library is excluded before association. The runtime result of associating across models is not described in the documentation and is not asserted here.
- Treating an image uploaded on the query side as making the image retrievable. The input is an image query and a library where no image index was generated at ingestion. Compare the ingestion index record with the query result, and observe whether the reference content is empty. An empty result is checked against the output path whose array length can be 0.
- Treating the minimum relevance score as a permission or entity filter. The input is the same question group under different relevance settings. Compare the referenced original text and scores before and after filtering, and observe whether unrelated business records can still be recalled. A relevance score does not perform authorization judgments.
How to verify the setup
- For the draft sentence "the deduction ratio during the 15 June activity period is 22%", input this sentence together with the contract and activity letter text. Compare whether the reference content includes Clause 3 and the activity letter's date range, and observe whether the referenced original text can support that ratio statement.
- For the handover request in the sample, input the request text. Compare whether the reference content covers both the activity letter's time range and the reconciliation statement's settlement basis, and observe whether the references are sufficient to check the adjustment target. If either item is missing from the references, record the check as not satisfied.
- For empty-result handling, input the paper input
[]. Compare how the downstream node handles a reference array of length 0, and observe whether the output path still runs. Do not claim that a fictional rule search necessarily returns[].
After any model or index adjustment, compare referenced original text using the same question group. Do not promise results based on vector scores alone.