What the documents look like
Central chain procurement and distribution work runs on collective procurement contracts, replenishment plans, delivery notes and store requisition forms. The supporting data comes from ERP purchasing, WMS, store POS and supplier delivery records. In cross-border operations the same batch of material frequently exists in several language versions: contract clauses, product names and delivery addresses may appear side by side in different languages. Material updates follow a daily and weekly rhythm. At retrieval time the different language versions of one document must be kept apart, so that a translation is not treated as evidence of the original.
These characteristics shape the retrieval task in a specific way. The document set is mixed in language and in type, and the same business object can be represented by a contract clause, a plan number and a delivery address in different languages. A retrieval step that returns a translated clause cannot by itself show what the original clause said, and a retrieval step that returns a plan number cannot by itself show that delivery was completed.
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 using different vector models cannot be associated together. Semantic retrieval supports cross-language understanding, so a Chinese question can match an English knowledge point, but its precision depends on the quality of model training and is affected by keywords and sentence completeness. Full-text retrieval suits exact queries such as numbers, models and proper nouns. Hybrid retrieval uses both vector and full-text recall, merges the results through RRF, and then applies ReRank for a second ordering. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The reference limit is measured in tokens and controls how much content a single search may cite. Retrieval itself is not authorization, and a form collecting values does not equal completed business verification or approval.
Hypothetical exercise sample: a chain supply chain center receives an excerpt from a collective procurement contract. The original text reads: "Party A issued replenishment plan number BP-2025-031 to Party B on 1 March 2025, agreeing to deliver in two batches to the East China distribution center." A draft sentence awaits judgment: "The replenishment plan completed full delivery on 1 March 2025." The review material contains only the contract excerpt above; no delivery note or receipt record is attached.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base containing the contract excerpt | First check the vector model identifier of each base, exclude those with a different identifier, then run retrieval |
| Search mode | Hybrid retrieval | Contract number BP-2025-031 is an exact query while the clause meaning needs vector recall; compare the candidates returned under each mode |
| Reference limit | Set to cover the full text of the contract excerpt | Base it on the actual length of the excerpt; compare truncated and complete returns |
| Minimum relevance | Start with a lower value to observe recall range | Label relevant and irrelevant samples before adjusting; compare candidates before and after filtering |
| Question optimization | Off for this trial | The draft sentence has a clear subject; compare whether the retrieval terms are rewritten when the setting is on and off |
| Result reranking | Not enabled for this trial | No rerank model is configured; its presence is not assumed |
Three common errors
- Treating "the contract agrees to deliver in two batches" as "full delivery completed." The input is the contract excerpt and the draft sentence; the comparison is the delivery status field. The contract excerpt states only an agreement and does not state that actual delivery was completed, so the draft sentence is unsupported.
- Treating minimum relevance as a permission filter. The input is a query involving an unauthorized store; the comparison is the returned content against the access permission list. A relevance threshold does not exclude unauthorized records, and permission must be verified separately.
- Associating knowledge bases that use different vector models. The input is two bases with different vector model identifiers; the comparison is the node association condition. The same vector model is the precondition for association, so bases whose identifiers do not match should be excluded and the actual observation recorded.
How to verify the setup
- Input the contract excerpt and the draft sentence "The replenishment plan completed full delivery on 1 March 2025." Compare whether the contract excerpt states that delivery was completed. The draft sentence is supported only when the excerpt explicitly records receipt or completion of delivery. This excerpt does not record it, so the judgment remains pending verification.
- Input the query "BP-2025-031 delivery batch." Compare whether the candidates returned by hybrid retrieval include the contract excerpt, and observe the number match and the semantically recalled fragment. A first-position match is not guaranteed.
- Input an empty-reference scenario, using the paper input [] to check whether the downstream node still runs. The cited content array may have a length of 0 and the output path can still continue. No claim is made that retrieval will necessarily return an empty array.
This page addresses only central chain procurement and distribution, cross-border material and multilingual retrieval, and knowledge base retrieval and recall. At operational handover, the authorization state of materials, operations and requests cannot be replaced by retrieval or by a form.