What the documents look like
Brand concession operations generate procurement and supplier records from several systems at once: mall POS, brand ERP, mall member campaigns and finance. A concession counter accepts unified mall collection and deduction-based settlement, so supplier delivery notes, joint-operation deduction contracts, mall campaign letters, sales receipts and monthly reconciliation statements circulate in parallel. Contract and campaign material tends to move on a monthly rhythm; delivery and receipt records move daily.
For document parsing and chunking, each record must keep the document identity and version that allow a later reader to trace a fragment back to its source. In a joint-operation deduction contract, the deduction percentage, the settlement cycle and the campaign letter's resource commitments belong to different parties and different periods. If chunking separates a condition from the content it governs, later verification loses its basis for judgement.
Constraints at the document chunking step
Chunking follows the structure of each document type. Contract material is best segmented by clause. A campaign letter is best segmented by campaign object, period and resource item. Receipts and reconciliation statements are best segmented at the level of the individual document. Original data and retrieval indexes stay distinct: one group of original data may map to multiple indexes, and adding indexes does not rewrite the original data.
Hypothetical paper exercise sample. A brand concession counter and a shopping mall hold a joint-operation deduction contract. The contract text contains three sentences: "This contract is valid from 1 January 2025 to 31 December 2025. The monthly sales deduction rate for the counterparty concession counter is 18%. The first party issues the previous month's reconciliation statement by the 15th of the following month." A separate supplier delivery note states: "Delivered 120 spring knitwear items on 5 March 2025; unit price not listed." The draft sentence to be assessed reads: "The deduction rate for this concession counter was adjusted to 20% in March 2025, and the supplier delivery note can serve as the basis for deduction settlement."
In this sample, the contract parties are the brand concession counter and the mall, the objects are the deduction rate and the reconciliation statement, the status is within the contract term, and the effective scope is limited to the 2025 year. The draft sentence treats an unrecorded rate adjustment and a delivery-note purpose as established facts. Retrieval and manual verification must handle these two claims separately.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid search | Contract clauses contain numbers and proper terms; hybrid search recalls by semantics and full text together, compared with semantic search alone for hit coverage |
| Reference limit | Hypothetical starting point of 2000 tokens | This value is a starting point for this sample only, compared with the length of a single contract clause, and is not an official default |
| Minimum relevance | Hypothetical starting point of 0.4 | Used to filter low-relevance fragments, compared with rerank scores to decide retention |
| Question optimization | Not enabled for this sample | The sample question has a clear subject; compare with the enabled run on elapsed time before deciding |
| Result rerank | Not enabled for this sample | Requires a rerank model enabled in settings; compare ordering against the run without rerank |
| Chunking basis | By clause and by individual document | Compared with fixed-length segmentation, checking whether a condition and the content it governs stay in the same segment |
The values in the table are hypothetical starting points for this exercise. They must be compared against other values for the same field before any conclusion, and they are not official defaults. RRF is the fusion process for hybrid retrieval results and is not configured as a standalone switch.
Three common errors
- Treating a successful upload as a completed retrieval update. After material changes, the actual training and index state must be checked; the upload action alone does not show that the index was rebuilt.
- Treating minimum relevance as a permission or entity filter. That value is a filtering threshold and does not exclude unrelated business or unauthorized records; retrieved results still require manual checking of party and period.
- Treating a pending review request as approved execution. If the rate adjustment in the sample draft sentence is only pending review, the deduction rate cannot be treated as changed, and the evidence gap must remain open.
How to verify the setup
- Input the three contract sentences and the draft sentence, then compare "the deduction rate was adjusted to 20%" in the draft against "the deduction rate is 18%" in the contract text. If retrieval returns only the 18% clause and no basis for an adjustment, the draft sentence is unsupported. If it returns a rate clause for another period or party, check whether the party and period match.
- Input the supplier delivery note text, then compare "unit price not listed" against the draft claim that the note "can serve as the basis for deduction settlement." If the note carries no amount and no field linking it to deduction, that purpose remains unverified and cannot be judged a settlement basis.
- Input an empty-result downstream scenario and check the downstream design against a paper input of []. If retrieval returns an empty array, check whether the downstream handles the empty result as such; do not claim that a fictional rule always returns []. Libraries built on different vector models cannot be associated together; check identifiers first, exclude non-matching ones, and record the actual observation rather than running a cross-model retrieval to compare missed recalls.
All three verification steps use the hypothetical sample above. Expected results must be compared and recorded in practice; no test is claimed to have been run. Semantic recall and relevance do not perform permission or entity isolation, and the same-vector-model requirement for retrieval libraries does not extend to external systems.