What the documents look like
Return and claim work handles reverse logistics and responsibility determination after export. The working materials are return agreements, claim letters, quality reports and return customs declarations. Sources include customer email, inspection reports, customs systems and insurer claim systems. Materials update per case rather than on a fixed schedule. Within one shipment, the agreement, letter, report and declaration must stay consistent on the responsible party, the batch, the dates and the current status.
Cross-border cases frequently hold the same material in Chinese and English side by side. Retrieval has to reach the same fact across languages. A retrieved passage is a candidate citation only. It does not establish responsibility and it does not grant authorization for any downstream business action.
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 built on different vector models cannot be associated together. Search operates on the whole knowledge base. A collection is a grouping device for management and does not change search behavior.
Hybrid search combines multiple recall paths through RRF. RRF is a fusion step inside the retrieval flow, not a switch to configure. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold and does not act as a permission, entity or business isolation control. The citation limit is measured in tokens and caps how much content one search may return. Question optimization adds one model call before retrieval and is used for coreference resolution and question expansion. Image retrieval requires image input together with the corresponding model capability. Reranking requires an available model and must be enabled in the settings. When a search returns nothing, the citation output is an array of length 0 and the downstream path still runs.
Hypothetical paper exercise sample: the source text for one shipment reads, "Return agreement TR-2024-018 states that the buyer submitted a return request on 12 May 2024, and the return reason is outer packaging damage. The attached quality report QR-2024-018 records that 30 units were sampled from the same batch, of which 4 units had outer packaging damage. Claim letter CL-2024-018 was issued by the buyer on 20 May 2024, and the claim amount is pending determination by the insurer." The draft sentence to be assessed is: "The return reason for this batch has been confirmed by the quality report as outer packaging damage, and the claim has been determined by the insurer." The draft touches the responsible party, the object, the status and the valid scope, and each item must be compared against the source text.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate one knowledge base containing return agreements, claim letters, quality reports and return customs declarations | First check the vector model identifier of each knowledge base, exclude any with a different identifier, and do not run retrieval across different models |
| Search mode | Use full-text search for identifier questions, semantic or hybrid search for natural language questions | Observe hits per test question, and do not assume one mode suits all materials |
| Citation limit | Hypothetical starting value of 3000 tokens | Test with materials of different lengths to see whether citations stay complete, and compare against larger and smaller values |
| Minimum relevance | Hypothetical starting value of 0.4 | Label relevant and irrelevant samples first, then observe how the value filters semantic retrieval or reranked results |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the responsible party or batch was altered |
| Result reranking | Use when a rerank model is available and enabled in the settings | Observe how rerank scores reorder candidate text, and compare against results without reranking |
Three common errors
- Treating a retrieval hit as authorization or business confirmation. Input the hypothetical source text and the draft sentence, compare "the claim amount is pending determination by the insurer" with "the claim has been determined by the insurer." If the draft states the claim is determined, that wording does not match the source status and turns a pending item into a completed one.
- Associating knowledge bases built on different vector models. Input the vector model identifiers of several knowledge bases, compare whether the identifiers match, and if they differ while association is still attempted, the node requirement for the same vector model is not met. The correct step is to check identifiers, exclude non-matching ones and record what was observed.
- Treating minimum relevance as a permission or entity filter. Input a query unrelated to return and claim work, compare the returned passages with the query subject, and if unrelated fragments are recalled, this does not show the setting can exclude unrelated business or unauthorized records. The value affects relevance filtering only.
How to verify the setup
- Input the hypothetical source text "Return agreement TR-2024-018 states that the buyer submitted a return request on 12 May 2024, and the return reason is outer packaging damage" together with the draft sentence "The return reason for this batch has been confirmed by the quality report as outer packaging damage," and compare where the return reason is recorded. If the source records the reason in the return agreement while the draft attributes it to the quality report, that attribution is unsupported and the source text must be checked again.
- Input the hypothetical source text "Claim letter CL-2024-018 was issued by the buyer on 20 May 2024, and the claim amount is pending determination by the insurer" together with the draft sentence "the claim has been determined by the insurer," and compare the claim status field. If the source says pending determination while the draft says determined, the two values are mutually exclusive in the same field and form a contradiction.
- Input a query with no corresponding material, compare the citation output length, and if the output is an array of length 0, record the empty result and check the downstream design. This observation does not show that every unrelated query returns empty, and it does not show that retrieval provides permission isolation.