What the documents look like
Returned-goods inspection handles reverse parcels and item grading. It sits on a different chain from in-warehouse picking and packing for new outbound orders. The main data sources are the WMS returns module, platform after-sales records, carrier return parcels, and the inspection bench. Daily documents include the return inbound receipt, inspection photos, the grade determination sheet, and the resale listing sheet. Data updates per item and per day.
After-sales questions ask whether a return has entered the warehouse, which material supports the inspection conclusion, and whether the item can be resold or refunded. Each answer must point back to one specific document, so citation sources and traceability become the central constraint for order fulfillment and after-sales question handling.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Its output is an array of cited content, and the length can be zero. When no content is found, this output path still completes. The cited content then passes to the answering step, where each answer sentence is checked against the cited source text.
Retrieval runs through semantic search, full-text search, hybrid search, RRF fusion, and ReRank reordering. Hybrid search merges multiple result paths through RRF. The reorder score can be used for filtering. Minimum relevance takes effect only in semantic search or when result reordering is used. It is a filtering threshold and does not perform permission or entity isolation. Retrieval itself is not authorization. Business verification, approval, database writes, and external synchronization each need their own downstream implementation.
Hypothetical paper exercise sample: a return inbound receipt states "Return R-2041, item is a folding chair, signed into warehouse on March 12, 2025, outer carton damaged, chair surface scratched." The inspection photo note for the same batch states "chair surface scratch approximately 3 cm, function normal." The grade determination sheet records "appearance defect, resalable." The draft sentence to be checked is: "This return has been determined as scrapped, and the refund has been approved." The scrap conclusion and refund status do not appear in either material.
Configuration and implementation checks
| Check item | Assumed trial starting point | Comparison object and reason |
|---|---|---|
| Associated knowledge bases | Select the four bases: return inbound receipt, inspection photo, grade determination sheet, resale listing sheet | Check whether they share the same vector model; different models cannot be associated together |
| Search mode | Start with hybrid search | Observe recall differences against semantic search and full-text search separately |
| Minimum relevance | Assumed starting value 0.5 | Compare with 0.3 and 0.7, observe changes in the number of cited items after filtering |
| Citation limit | Assumed starting value 2000 tokens | Compare with 1000 and 4000, observe the length of cited content |
| Result reordering | Enabled | Compare sort positions against the disabled state, confirm the model is configured |
| Citation handoff | Pass the citation array to the answering node | Check each answer sentence against the cited source text |
The starting values are set only for this paper exercise to establish a comparable observation baseline.
Three common errors
- Treating a retrieval hit as business authorization. Input "R-2041 scrap refund" and retrieve the grade determination sheet. Compare the retrieval result with the determination sheet text. The sheet only states "appearance defect, resalable." The scrap and refund conclusions are not supported. Retrieval does not produce approval or execution.
- Treating a similarity score as a permission or entity filter. Input the same question and adjust minimum relevance. Compare the number of returned cited items. The score changes recall scope only. It does not distinguish authorized parties from unrelated records.
- Writing a pending-review status as completed. Input the draft sentence "refund has been approved" and compare it with the platform after-sales record. If the record shows only "pending review," the sentence is not supported. Pending review does not equal approval.
How to verify the setup
- Input the return inbound receipt and inspection photo text above, and ask "can R-2041 be resold." Compare the answer sentence with the grade determination sheet text. If the answer cites only "appearance defect, resalable" and marks the source, the condition is met. If a scrap or refund conclusion appears, the condition is not met.
- Input the draft sentence "this return has been determined as scrapped, and the refund has been approved" and compare it with the three material texts. None of the materials state scrap or refund. The answer should indicate the claim is unsupported and preserve the evidence gap. This meets the condition. Directly confirming the sentence does not meet the condition.
- Input an empty citation array [] with one normal question, and compare the downstream answering node behavior. If the downstream handles the empty result branch and indicates no content was retrieved, the condition is met. If the downstream errors out or fabricates a source, the condition is not met. For knowledge bases with different vector models, first check the vector model identifier and exclude those that do not match. Record actual observations. Do not infer runtime results for retrieval that was not executed.