What the documents look like
Supplier entry and settlement records centre on purchase contracts, delivery acceptance notes, rebate agreements and statements of account. These materials originate in ERP procurement, supplier portals, POS sales and finance systems. Settlement obligations usually take the form of purchase-and-sale or consignment payables, and a resident on-site operation does not necessarily exist. Records update weekly or monthly. A single statement of account may carry Chinese contract clauses alongside foreign-language portal entries, so retrieval has to reach the same business fact across languages.
The eight-dimensional brief for this category identifies verification of accepted quantities, calculation of rebate deductions and handling of return differences as the most time-consuming steps. Each of these steps depends on matching a document reference, a date range and a quantity scope across languages. Retrieval quality therefore affects whether staff can locate the correct clause or line item before any reconciliation decision is made.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Semantic retrieval relies on vector distance and supports cross-language understanding, which suits natural-language questions. Full-text retrieval relies on keyword hits, which suits identifiers, model numbers and proper nouns. Hybrid retrieval combines both paths through RRF, and ReRank performs a second ordering of candidate text. The minimum relevance score takes effect only under semantic retrieval or when result reordering is used. The reference limit is measured in tokens. Empty output is returned as an array whose length can be zero, so the downstream path still runs when no content is found. Retrieval does not perform permission or entity isolation, and a retrieval hit does not constitute authorization.
Assumed paper exercise sample. A supplier portal record states: "Delivery acceptance note number DN-2024-0871, acceptance date 12 June 2024, accepted quantity 480 boxes, status pending finance review." A statement of account for the same period states: "DN-2024-0871 corresponds to a settled quantity of 480 boxes; the rebate deduction follows Annex 2 of the contract." A draft sentence awaiting judgement reads: "DN-2024-0871 has completed settlement and been paid." The materials under review do not state a payment status or a settlement completion confirmation, so this draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the Chinese contract base and the foreign-language portal base that share one vector model | Confirm the vector model identifier matches before associating; exclude any base whose identifier differs |
| Search mode | Test the same question separately under semantic retrieval and full-text retrieval | Compare hit differences between the two modes for identifier-type and natural-language questions |
| Reference limit | Assumed starting point set to cover one complete statement of account in tokens | Test with materials of different lengths to see whether content is truncated, then adjust |
| Minimum relevance score | Assumed starting point at a mid value | Label relevant and irrelevant samples first, then compare filtering effects |
| Question optimisation | Enable only for multi-turn follow-ups or short questions | Compare the original question with the optimised question and check whether the subject was altered |
| Result reordering | Enable when candidate results are numerous | Compare ordering before and after reordering and confirm the question has a complete subject and predicate |
Three common errors
- Treating a retrieval hit as an authorization basis. Input is the draft sentence "DN-2024-0871 has completed settlement and been paid." The comparison is against the source materials, which state only an accepted quantity and a pending-review status. The observable result is that payment status is not stated, so the draft sentence is unsupported.
- Forcing knowledge bases with different vector models into one association. Input is two bases whose vector model identifiers differ. The comparison is against the node requirement that associated bases share the same vector model. The observable result is that the mismatched base is excluded by identifier check; no cross-model retrieval run is performed to compare missed recalls.
- Using the minimum relevance score as a permission or entity filter. Input is an unrelated business record with a high similarity score. The comparison is against the documented scope of the minimum relevance score, which filters low-relevance results only. The observable result is that this threshold does not exclude unauthorized records, and retrieval does not provide permission isolation.
How to verify the setup
- Input the portal record and the statement of account from the assumed sample. Compare the accepted quantity and the settled quantity for DN-2024-0871. The condition is satisfied when both locations state 480 boxes.
- Input the draft sentence "DN-2024-0871 has completed settlement and been paid." Compare whether the materials state a payment status or a settlement completion confirmation. When neither is stated, the sentence is judged unsupported and the evidence gap is retained.
- Input a deliberately constructed empty result array. Compare whether the downstream path continues to run with reference content whose length can be zero. The condition is satisfied when the downstream design can accept an empty array; no claim is made that a fictional rule will always return an empty array.