What the documents look like
Trade returns and claims work handles reverse logistics and liability determination after goods have been exported. The materials come from customer emails, quality inspection reports, customs systems, and insurer claim systems. The documents handled day to day include return agreements, claim letters, quality reports, and return customs declarations. Each case is different, and updates are irregular rather than scheduled.
For product data and selection questions, this matters because an answer has to trace back to the original document line by line. The knowledge base search node returns cited content as an array, and that array may have a length of zero. A zero-length result still lets the workflow path continue; it does not mean the workflow has stopped with an error. Before any search is configured, the selected knowledge bases and the document versions in them need to be checked. Only knowledge bases that share the same vector model can be associated together.
Constraints at the citation tracing step
The search node outputs cited content as an array, and downstream nodes must handle that array shape, including the empty case. A similarity score reflects relevance only. It does not act as a permission filter or an entity filter, and a high score does not mean a record has been authorized for use. Hybrid search merges multiple recall paths through RRF, result reranking runs when a rerank model is available and the setting is enabled, and image retrieval requires the corresponding image input and model capability. The overall retrieval flow lists several optional paths; not every path runs on every request.
Hypothetical exercise sample: the source text to check against is "On March 10, 2025, return agreement TN-2025-031 states: the buyer accepted batch B2401 on February 20, 2025, and applied for a full-batch return because the appearance dimensions were out of tolerance. The return reason is pending confirmation by the quality inspection report. Quality inspection report QR-2025-118 has the status pending review." The draft sentence to be judged is "The return reason for this batch has been confirmed as quality failure, and the claim letter can be submitted directly." The subject, object, status, and validity scope needed for the judgment all appear in the source text. The quality inspection report is in pending review status, and the return reason has not been confirmed.
Configuration and implementation checks
| Configuration or check item | Setting relevant to this page | How to check |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases that share the same vector model | Check the selected base identifiers and document versions, and exclude any base with a different vector model |
| Search mode | Semantic search, full-text search, or hybrid search | Choose by question type; identifier-style questions can use full-text or hybrid |
| Citation limit | Maximum tokens cited per search | For this hypothetical exercise the starting value is 3000 tokens, chosen because the sample material is short, then compared with the displayed value and adjusted |
| Minimum relevance | A 0-1 value that takes effect only under semantic search or result reranking | For this hypothetical exercise the starting value is 0.3, compared against retrieval result scores |
| Result reranking | Requires a rerank model and runs when enabled | Check model availability and enablement state |
| Cited content output | Array format, length may be zero | Use the paper input [] to check the downstream handling design |
Three common errors
- Treating a pending-review quality inspection report as a confirmed conclusion. The input is the sample source text and the draft sentence. Compare "pending review" with "confirmed as quality failure." If the draft sentence appears, it is marked as unsupported.
- Treating a similarity score as a permission filter. The input is the minimum relevance setting and the retrieval results. Compare the score with the authorization scope. A high score does not mean the record has been authorized.
- Associating knowledge bases that use different vector models. The input is the list of selected base identifiers. Compare the vector model identifiers. If they differ, exclude that base and record the observation.
How to verify the setup
- Input the sample source text and the draft sentence, then compare the return reason field status. The draft sentence is marked as unsupported and the evidence gap is preserved. This meets the condition.
- Input the paper value
[]as the cited content output, then compare whether the downstream node continues to execute. The path continues without an error. This meets the condition. - Input the list of selected knowledge base identifiers, then compare the vector model identifier of each base. All must match before they can be associated. If any differ, exclude that base and record the actual observation. This meets the condition.
These checks describe expected behavior. They do not claim that a test, deployment, approval, or refund has taken place. Retrieval is not authorization, and a pending request is not an approval. Where the documents do not state a field, the value remains unknown.