What the documents look like
Trade return and claim work for this category turns on four document types: the return agreement, the claim letter, the quality report, and the return customs declaration. Materials are assembled case by case and updated irregularly, so a product information and selection question-and-answer setup has to work across documents that arrive at different times and carry different levels of detail.
The questions users bring to such a knowledge base mix two shapes. One shape carries exact strings: a model number, a customs declaration number, a batch identifier. The other shape is ordinary natural language, such as asking which specification a shipment was declared under. Exact identifiers suit full-text retrieval, while natural-language phrasing depends on semantic recall. Both shapes coexist inside one knowledge base, and that coexistence is the main data characteristic at this stage.
Retrieval returns quoted source text only. Whether a quoted passage may be submitted externally, and whether any submission has been authorized, are separate questions that require manual review and a downstream business process. The vector index does not decide those questions.
Constraints at the vector indexes 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. When checking this, inspect the model identifier first, exclude and record any knowledge base whose identifier does not match, and do not run a cross-model search to compare missed recalls.
A single original data block can map to multiple indexes. Index content and original content are checked separately. An index that is complete and focused supports retrieval coverage, and the trade-off between coverage and index length is weighed against the samples on this page. Image retrieval additionally depends on whether an image description index or an image vector index was generated at ingestion time. Uploading an image on the query side does not by itself make that image retrievable.
Semantic retrieval, full-text retrieval, hybrid retrieval, and result reranking are optional paths. A single request does not necessarily execute all of them. Image retrieval requires the corresponding image input and model capability. Reranking requires a rerank model and must be enabled according to the configuration. RRF is the fusion process applied to multiple result paths after hybrid retrieval; it is not a switch that can be turned on. The minimum relevance score is a filter threshold that takes effect only under semantic retrieval or when result reranking is used, and it does not perform permission or entity isolation.
Configuration and implementation checks
Hypothetical drill sample: an export lighting shipment is returned. The source materials read as follows. The return agreement states that the return reason is housing cracks, covering model L-220, quantity 400 units, with an agreement signing date of March 10, 2025. The quality report states that 6 of 20 sampled units showed housing cracks, report number QR-2025-031, issued March 12, 2025. The draft sentence to be assessed is: "The L-220 lighting shipment was returned due to housing cracks, and the claim letter has been submitted to the insurance company." The subject, object, status, and date scope are limited to these two materials.
| Configuration item | Value or check item on this page | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only return and claim knowledge bases sharing one vector model | Check the model identifier first; exclude and record knowledge bases with a different identifier |
| Search mode | Hypothetical trial starting point for this drill is hybrid retrieval | Compare quoted source text against separate semantic retrieval and separate full-text retrieval on the same question set |
| Minimum relevance score | Hypothetical trial starting point for this drill is 0.5 | This value takes effect only under semantic retrieval or reranking; compare against whether quoted text contains the target model number |
| Reference limit | Set in tokens; hypothetical trial starting point for this drill is 2000 tokens | Compare against whether returned references cover the key fields of both materials |
| Index check | Check item by item whether model number, report number, and dates entered the index | Compare character by character against the source text; record missing items as pending |
| Result reranking | Enabled or not according to configuration | After enabling, compare rerank scores against the relevance of quoted source text |
Three common errors
- Associating knowledge bases that use different vector models. Input: two knowledge bases with different model identifiers. Comparison item: whether the model identifiers match. Observable result: when identifiers differ, exclude and record them, and do not continue retrieval.
- Treating the minimum relevance score as a permission or entity filter. Input: one low-relevance record that belongs to a different case. Comparison item: whether the returned references include that record. Observable result: the threshold affects relevance ordering only, and record ownership still requires manual review.
- Treating an image index as automatically present on the query side. Input: one ingested image with no image index generated, plus one image query. Comparison item: whether a corresponding image description index or image vector index exists in the knowledge base. Observable result: with no index, that image does not enter recall.
How to verify the setup
- Input the two source materials above and the draft sentence. Compare whether "the claim letter has been submitted to the insurance company" appears in the materials. Observable result: the materials do not state this fact, so the sentence is judged unsupported, and the evidence gap is retained.
- Input model number L-220 and report number QR-2025-031. Compare whether the full-text retrieval path hits the corresponding source text. Observable result: a hit passes; a miss records an index gap for that field and does not by itself assert that retrieval is ineffective.
- Input one empty-result query. Compare whether the downstream accepts an array of length 0 as reference content. Observable result: the path completing passes; no claim is made that any particular retrieval rule necessarily returns an empty array.