What the documents look like
Central procurement and central distribution materials follow head-office purchasing and outbound delivery flows. The main data sources are ERP procurement records, WMS records, store POS data, and supplier delivery data. Everyday documents include centralized procurement contracts, replenishment plans, delivery notes, and store requisition forms. Operating reports and exception reviews compare the same entity, the same period, and the same field across these materials. A typical comparison asks whether the quantity a store requisitioned on a given day matches the quantity actually dispatched on the delivery note for that store.
Material updates occur daily or per document. Retrieval configuration only determines how reference content is recalled. It does not change how the underlying business data is updated. A knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. This constraint matters because operating reviews often need to pull contract text, replenishment plans, delivery notes, and requisition forms into one reference set, and the linkage decision must respect the same-vector-model requirement before any search runs.
Constraints at the retrieval step
Semantic retrieval, full-text retrieval, and hybrid retrieval serve different material. Document numbers, model codes, and proper nouns suit full-text or hybrid retrieval. Natural-language questions suit semantic or hybrid retrieval. Hybrid retrieval combines vector and full-text recall through RRF, which is a fusion process for multiple result paths, not a toggle that can be switched on. The minimum relevance value is a 0-1 filter threshold. It takes effect only under semantic retrieval or when result re-ranking is used. It is a filter threshold, not a final sorting rule, and it does not exclude unrelated businesses or unauthorized records. The reference cap is measured in tokens, not in a top-k item count. Question optimization performs coreference resolution and question expansion before retrieval and adds one model call. Re-ranking requires a re-rank model and must be enabled through its settings. Image retrieval requires image input and the corresponding model capability. The overall retrieval flow lists optional paths; not every request executes all of them.
Retrieval itself does not carry authorization or entity isolation. Whether recalled content supports a business judgment still requires manual comparison.
Hypothetical drill sample: the material text reads, "The East China Distribution Center delivery note PD-0603 dated June 2024 records an actual dispatch to store A17 of 120 cases of a specified product specification; store requisition form YH-0603 records a requisition of 100 cases." The draft sentence to be assessed is, "Store A17's requisition and actual dispatch quantities on June 3 are consistent." The check compares the quantity fields of the two documents. Entity, date, and product specification are the same; the quantities are 100 and 120. The draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the same-vector-model knowledge base containing the delivery note and requisition form above | Verify the vector model identifier first, exclude knowledge bases on different models, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | The draft sentence contains a document number and a quantity; keyword matching and semantic recall are used together and compared against pure semantic and pure full-text results |
| Reference cap | Set to a token amount that covers the full text of both documents | Test content completeness with materials of different lengths, then adjust |
| Minimum relevance | No threshold set at the start; observe recalled content first | This value takes effect only under semantic retrieval or result re-ranking; label relevant and irrelevant samples first |
| Question optimization | Off for this trial | The draft sentence has a clear subject; compare whether enabling it alters the subject |
| Result re-ranking | Not enabled for this trial | Requires a re-rank model enabled through its settings; with few candidates, observe raw recall first |
Three common errors
- Treating a retrieval hit as authorization or business confirmation. Returning the delivery note and requisition form text does not mean a quantity difference has been approved, a supplier claim has been established, or inventory has been adjusted. Those are business actions that require separate downstream system implementation.
- Treating minimum relevance as an entity or permission filter. The value is a 0-1 filter threshold and does not exclude unrelated businesses or unauthorized records. Cross-store and cross-supplier isolation still has to be implemented outside the retrieval scope.
- Forcing knowledge bases on different vector models into one linkage. Verify the identifier first, exclude those that do not match, and record what is observed. Do not infer missed recall or redundant results from a forced linkage.
How to verify the setup
- Enter the two document texts and the draft sentence from the hypothetical sample. Compare the quantity field in the draft sentence against the source values 100 and 120. The condition is met when the result is "not supported" and the reference content contains both documents.
- Enter a query containing only the document number "PD-0603." Compare whether the recalled content includes the delivery note text and whether the reference cap fully covers its quantity and entity fields. The condition is met when the reference is complete.
- Use a paper input of an empty reference array to check the downstream design. Compare whether the downstream handles the empty-result branch. Do not claim that a fictional rule retrieval necessarily returns an empty array.
This page addresses only central chain procurement, operating reports and exception reviews, and knowledge base retrieval and recall. Retrieval or a form does not constitute authorization to execute a business action. Whether recalled content can support a specific business judgment remains subject to manual verification.