What the documents look like
Reverse returned-goods inspection handles returned parcels and merchandise grade decisions. It does not handle outbound picking for new orders. The work draws on the WMS returns module, platform after-sales records, carrier return parcels and the inspection station. Each returned item produces a return inbound note at receiving. The inspection station keeps inspection photographs. A grade decision table records the condition conclusion for the item. A re-sale listing note carries items that can be sold again. For the procurement and supplier reference task, supplier qualifications, purchase batches and return liability attribution are organised into a knowledge base for retrieval.
These documents update per item and per day. Retrieval works only against material already in the knowledge base. This page does not infer any batch synchronisation mechanism from the update frequency, and the update frequency does not change the retrieval step described here.
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. When candidate bases are reviewed, compare the vector model identifier first, exclude those that do not match, and record what was actually observed. Do not run a cross-model retrieval to compare missed recalls.
Semantic retrieval relies on vector distance. Full-text retrieval relies on keyword hits. Hybrid retrieval merges the two paths through RRF. RRF is a fusion process, not a switch that is turned on. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold, not a final ordering rule, and it does not carry permission or entity isolation. The reference cap is measured in tokens. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and must be enabled by configuration. The optional paths shown in the overall flow diagram do not mean every request executes all of them. Retrieval output is an array and its length can be 0, so the empty-result path still proceeds. A downstream design can be checked against an explicit paper input of [].
Hypothetical paper exercise sample: the material text supplied for checking reads, "Return inbound note RT-2024-0417 states that batch B-77 supplied by Supplier A was returned to the warehouse on 17 April 2024, quantity 12 items; the grade decision table conclusion is surface scratches, assigned grade C; the re-sale listing note has not yet been generated." The handover request to be judged reads, "Based on the above material, confirm that batch B-77 can be re-sold and synchronise it to the procurement system." Whether grade C permits re-sale, and whether procurement system synchronisation has been implemented, are not stated in the material.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the procurement and supplier reference base | Matches the primary task field; compare the vector model identifier of each base for sameness |
| Search mode | Run semantic retrieval and full-text retrieval separately | Semantic suits natural-language description, full text suits exact batch numbers; compare the two recall sets |
| Reference cap | Start at a token amount that covers one complete grade decision table | Chunk lengths differ widely; compare recall completeness against longer and shorter actual material |
| Minimum relevance | Start at 0.5 | First label relevant and irrelevant samples, then compare filtering against adjacent values |
| Question optimisation | Enable for follow-up questions that contain references | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous | Rerank scores run 0 to 1 and filter by score; compare against ordering without reranking |
Three common errors
- Treating a retrieval hit as authorised execution. The input is the handover request "confirm that batch B-77 can be re-sold and synchronise it to the procurement system." Compare it with the material text. The material states grade C and that the re-sale listing note has not been generated. It does not state that grade C permits re-sale, and it does not state that procurement system synchronisation has been implemented. The observable result: retrieval returns grade decision table content, while the business action in the request remains unsupported and needs a separate system implementation.
- Treating minimum relevance as a permission filter. The input is a query about an unrelated supplier. Compare the returned content with the query subject. Minimum relevance filters by score only; it does not isolate by entity. The observable result: low-scoring content is filtered out, while content of the same score is not excluded because it belongs to a different supplier.
- Associating knowledge bases across different vector models. The input is two candidate bases. Compare their vector model identifiers. A difference in identifier means exclusion. Record the observation and do not infer the runtime result of a joint association.
How to verify the setup
- Use the handover request from the hypothetical sample as input. Compare the retrieved content with the material text. The request is supported only when the returned content explicitly states that grade C permits re-sale and that procurement system synchronisation has been implemented. Otherwise record it as unsupported.
- Use batch number B-77 as input for full-text retrieval and a natural-language description for semantic retrieval. Compare the two recall sets. Observe whether the number-type query is hit. Do not expect a first-position hit or complete recall.
- Use the paper input [] to check the empty-result path. Compare whether downstream nodes still execute. Observe whether the path proceeds when the reference content length is 0. Do not claim that a particular rule retrieval always returns [].