What the documents look like
Reverse returned-goods inspection handles returned parcels and product grades. The working materials are the return warehouse receipt, inspection photographs, the grade determination table, and the resale listing sheet. Updates occur per item and per day. In a cross-border setting, one batch of returned items may arrive from storefronts in different languages, so after-sales tickets, carrier return labels, and inspection notes mix several languages. Retrieval has to match the same product or the same return reason across those languages.
Inspection photographs and the grade determination table drive the resale-or-scrap decision. If a retrieval result cites the determination clause for the wrong grade, inventory disposition is directly affected. Materials are item-grained. One return number, one product code, and one batch can be spread across several documents, so recall has to cover both exact identifier queries and natural-language descriptions.
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 on different vector models cannot be associated together. Verify the model identifier first, exclude those that do not match, and record what was actually observed.
Semantic retrieval relies on vector distance and suits multilingual natural-language questions. Full-text retrieval relies on keyword hits and suits exact queries such as return numbers, product codes, and error codes. Hybrid retrieval uses both recall paths and fuses them through RRF; RRF is a fusion process, not a standalone switch. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold. It does not carry out permission or entity isolation, and it does not exclude unrelated business records. The reference limit is measured in tokens and controls how much content a single search may cite.
Hypothetical paper exercise sample: an after-sales warehouse receives one returned item. The source material reads: "Return warehouse receipt RV-2026-0417, product code SKU-88231, return reason field states 'outer packaging damaged, product body shows no damage', inspection bench note states 'housing scratch approximately 2 cm, functional test passed'." The draft sentence to be assessed reads: "This returned item is determined resalable, grade A, and a resale listing sheet may be generated directly." The source contains the subject (after-sales warehouse), the object (RV-2026-0417 / SKU-88231), the states (outer packaging damaged, body undamaged, housing scratch, function passed), and the effective scope (this receipt, this product). The draft sentence adds grade A and a direct-listing conclusion that the source does not state. It is unsupported.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases using the same vector model | Return warehouse receipts, inspection photographs, grade determination tables, and resale listing sheets are ingested separately; verify the vector model identifier matches |
| Search mode | Compare semantic, full-text, and hybrid retrieval per test question | Identifier and code questions use full-text or hybrid; multilingual natural-language questions use semantic or hybrid |
| Reference limit | Unit is tokens; for this exercise a starting value of 2000 is assumed | This starting value is chosen to compare whether short and long return materials are cited in full, against 1000 and 4000 |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking | Label samples related and unrelated to RV-2026-0417 first, then adjust the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the multi-turn follow-up "can this one be resold" before and after optimization, checking whether the subject still points to RV-2026-0417 |
| Result reranking | Requires a ReRank model and is enabled per settings | When candidates are numerous, rerank inspection clauses and check the score against question completeness |
Three common errors
- Treating "grade A, a resale listing sheet may be generated directly" in the draft sentence as already stated in the source. The source records only outer packaging damage, an undamaged body, a housing scratch, and a passed functional test; the grade and the listing conclusion do not appear. To check, input the draft sentence and the source, compare whether the grade field and the listing action exist in the source, and record that a field absent from the source is unsupported.
- Treating minimum relevance as a permission or entity filter. The value takes effect only under semantic retrieval or result reranking and filters low-relevance results; it cannot exclude unrelated business records or unauthorized records. To check, input the same query at different relevance thresholds, compare whether the returned content still contains unrelated return receipts, and confirm that changing the threshold does not change the authorization scope.
- Describing RRF as a standalone switch that can be turned on, or mistaking the unit of the reference limit for item count. RRF is the fusion process for hybrid retrieval results, and the reference limit is measured in tokens. To check, inspect whether an RRF switch exists in the interface, inspect the unit of the reference limit value, and observe whether the configuration names match the documented description.
How to verify the setup
- Input the hypothetical source text "Return warehouse receipt RV-2026-0417, product code SKU-88231, return reason field states 'outer packaging damaged, product body shows no damage', inspection bench note states 'housing scratch approximately 2 cm, functional test passed'" together with the draft sentence "This returned item is determined resalable, grade A, and a resale listing sheet may be generated directly." Compare whether the grade and the listing conclusion in the draft sentence have corresponding fields in the source. The source does not state the grade or the listing authorization, so the result is unsupported and requires the grade determination table source before further checking.
- Input the query "inspection conclusion for RV-2026-0417" and run it under semantic, full-text, and hybrid retrieval respectively. Compare whether the cited content returned by each path contains the inspection note for that receipt number. Full-text or hybrid retrieval is expected to hit the identifier; the semantic result depends on the vector model and the language distribution of the material, with no guarantee of a first-position hit or complete recall.
- Input a paper-based empty-result continuation case: assume a query returns cited content of
[], and check whether the downstream flow proceeds through the output path as documented. This empty array is a hypothetical input used to check how the downstream design handles zero citations; it does not claim that a fictional rule search always returns an empty array. The retrieval step itself does not perform authorization, approval, or database writes; resale listing requires a separately connected system implementation.