What the documents look like
Central chain procurement retrieval material centres on headquarters purchasing and distribution. Sources named in the brief are ERP procurement, WMS, store POS and supplier delivery data. The documents handled day to day are centralised procurement contracts, replenishment plans, distribution orders and store requisition orders, with updates at daily or weekly frequency.
For order fulfilment and after-sales question answering, these four document types are used to locate the subject, object, status and validity scope of a question. Retrieval returns cited content only. Authorization and execution sit downstream and require their own implementation.
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. At retrieval time, the vector model identifier is checked first; non-matching knowledge bases are excluded and the actual observation is recorded. A cross-model retrieval is not run to compare missed recalls.
Semantic retrieval, full-text retrieval and hybrid retrieval each suit different material. Queries built around identifiers and model numbers suit full-text or hybrid retrieval; natural-language questions suit semantic or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process, not a toggle. The citation limit is measured in tokens. The minimum relevance value takes effect only during semantic retrieval or when result reranking is used. Question optimisation performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and is enabled according to settings. Cited content is output as an array whose length may be zero; an empty result still passes through the output link. Retrieval and similarity scores do not perform permission or entity isolation.
Hypothetical drill sample: the source text reads "On 1 March 2025, Supply Centre A issued distribution order PS-2025-0301 to Distribution Centre B, requiring delivery to Store C by 5 March, with status Issued." and "This distribution order corresponds to store requisition order YH-2025-0228, requisition quantity 200 cases, with status Awaiting distribution centre confirmation." The draft sentence to be assessed is "Store C received the goods on 5 March, and any after-sales discrepancy is borne by the supplier." The source text records only issuance and pending confirmation. It does not record receipt, arrival or allocation of responsibility, so the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Centralised procurement contract base and distribution order base; check the vector model identifier first | They may be associated together only when the identifiers match; a mismatch excludes the knowledge base |
| Search mode | Try semantic, full-text and hybrid separately on the sample question | Compare whether the cited content from each path covers the order number and status field |
| Citation limit | Start at a token amount covering both source passages | Test content completeness with material of differing lengths, then adjust the limit |
| Minimum relevance | Mark relevant and irrelevant samples first, then set a threshold | Takes effect only during semantic retrieval or result reranking; compare before and after filtering |
| Question optimisation | Enable for short questions such as "that order on 5 March" | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous and check the score | Reranking requires a model and is enabled according to settings; compare the ordering change |
Three common errors
- Treating a retrieval hit as authorization. In the sample, the distribution order status is Issued and the requisition order status is Awaiting distribution centre confirmation. Retrieving these two source passages does not mean confirmation is complete, and does not mean receipt or allocation of responsibility is established.
- Treating similarity as a permission or entity filter. The minimum relevance value is a filtering threshold, not a final ordering rule, and cannot be used to exclude unrelated businesses or unauthorized records.
- Treating optional paths as executed on every request. Image retrieval requires an image input and model capability. Reranking requires a model and is enabled according to settings. When the conditions are not met, that path is not executed.
How to verify the setup
- Input the sample source text and the draft sentence, and compare whether the draft sentence exceeds the issuance and pending-confirmation status recorded in the source text. If it is judged unsupported and the evidence gaps for receipt and allocation of responsibility are retained, the condition is met. If it is judged as received or as supplier-borne, the condition is not met.
- Input order number PS-2025-0301, and compare whether the cited content from full-text retrieval and hybrid retrieval contains that order number and the status field. If the cited content contains the order number and the Issued status, the condition is met. If the cited content lacks that field, record the gap.
- Input an empty array [], and compare whether the downstream design handles a citation length of zero. If the downstream link passes through and produces no business action, the condition is met. If the downstream treats an empty result as confirmed, the condition is not met.