What the documents look like
Returns refurbishment and resale retrieval draws on overseas warehouse WMS records, TikTok return center data, Amazon return reports, and product inspection photos. Daily documents include return inspection sheets, refurbishment records, scrap declarations, and relisting checklists. These materials span Chinese, English, and local site languages, and the inspection conclusion, refurbishment action, and relisting decision for one return batch sit in separate documents. A search runs against the whole knowledge base, while collections only classify data and do not affect search results.
The vector model determines the cross-language reach of semantic recall. Knowledge bases linked for one search must share the same vector model, so the actual model and its input types are checked first. One source data block can map to several indexes, and the index content is reviewed separately from the source text. Index content is built from complete and focused information, and retrieval coverage against index length is weighed with samples. Whether an image can be retrieved depends on whether an image description or image vector index was generated at ingestion time; uploading an image on the query side is not a sufficient condition.
Constraints at the vector indexes step
A hypothetical drill sample is marked here. One overseas warehouse return inspection record reads: "Inspection date 2025-03-12, batch A-0312, product Bluetooth speaker, cosmetic scratch grade B, function test passed, recommend refurbishment then relisting." The scrap declaration for the same batch reads: "Batch A-0312, Bluetooth speaker, mainboard moisture damage, judged scrap, does not enter relisting." A draft sentence awaiting review reads: "Batch A-0312 Bluetooth speaker passed inspection, relisting has been arranged."
The draft sentence and the scrap declaration take mutually exclusive values on batch, product, and status fields: the inspection sheet says "recommend refurbishment then relisting," and the scrap declaration says "judged scrap, does not enter relisting." On the authorization boundary, the scrap declaration is signed by the inspection role, and relisting requires separate confirmation. Pending review does not show that operating permission is absent, and signing, reaching agreement, and contract effectiveness are three distinct facts. The retrieval step only recalls source text and does not carry out authorization.
Mixed retrieval combines vector recall and full-text recall through the RRF formula, and the merged range is large, so a rerank model is usually used to reorder and filter by its score. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result reranking is used, and it is not a final ordering rule. The reference cap limits how many tokens of content are cited per search; the document explains that token-based control is used instead of top k because chunk lengths differ widely in mixed knowledge bases.
Configuration and implementation checks
| Configuration item | Setting or check for this page | Link to the sample |
|---|---|---|
| Linked knowledge bases | Select bases with the same vector model; verify the model identifier and input types first | Inspection sheet and scrap declaration must be linkable under one model |
| Search mode | Semantic retrieval, full-text retrieval, or mixed retrieval | Batch number A-0312 suits full-text or mixed retrieval |
| Index content | One source data block can map to several indexes; review index and source text separately | Inspection conclusion and scrap conclusion each get their own index |
| Image index | Confirm whether an image description or image vector was generated at ingestion | The index status of inspection photos needs verification |
| Reference cap | Each search cites at most n tokens of content | The value is a hypothetical starting point for this drill and is compared against the same question group |
| Minimum relevance | A 0-1 value that takes effect only in semantic retrieval or result reranking | This threshold is a filter condition and does not isolate permissions |
The reference cap value used in the drill is a hypothetical starting point chosen to keep the cited content near one inspection record and one scrap declaration, and it is compared against the same question group run with a higher and a lower value. It is not an official default. The minimum relevance value in the drill is likewise a hypothetical starting point, chosen so that clearly unrelated content is filtered while both source records remain in the candidate set, and it is compared against the same question group at a higher and a lower value.
Three common errors
- Treating retrieval recall as authorization execution. The input is the draft sentence "Batch A-0312 Bluetooth speaker passed inspection, relisting has been arranged." Compare the batch, product, and status fields of the inspection sheet and the scrap declaration, and observe whether the draft sentence contradicts the scrap declaration's "judged scrap, does not enter relisting." A retrieved source text does not show that relisting has been approved.
- Treating minimum relevance as an entity or permission filter. The input is a cross-batch query. Compare whether recalled results include batches other than A-0312, and observe whether unrelated batch content appears. This threshold is a filter condition and cannot exclude unauthorized records.
- Treating query-side image upload as proof that an ingestion index exists. The input is an inspection photo query. Compare whether an image description or image vector was generated at ingestion, and observe whether a pure image query can match. When no index was generated, weak image-to-image results are expected behavior.
How to verify the setup
- Input the inspection sheet text "Inspection date 2025-03-12, batch A-0312, product Bluetooth speaker, cosmetic scratch grade B, function test passed, recommend refurbishment then relisting" and the draft sentence "Batch A-0312 Bluetooth speaker passed inspection, relisting has been arranged." Compare the status fields of the two, and observe whether the draft sentence turns "recommend refurbishment then relisting" into a completed relisting. The condition is met when the draft sentence does not match the source wording, which is judged unsupported.
- Input the scrap declaration text "Batch A-0312, Bluetooth speaker, mainboard moisture damage, judged scrap, does not enter relisting." Compare the same batch, same product, and same status field, and observe whether the two source texts take mutually exclusive values. The condition is met when the same field shows mutually exclusive values, which is judged a contradiction and needs manual verification.
- Input an explicit paper empty result []. Compare whether the downstream chain still runs, and observe whether the output is a reference array of length 0. The condition is met when the output chain runs and the reference is empty; no claim is made that a fictional rule search necessarily returns [].
When a model or index changes, compare cited source text on the same question group, and do not promise an effect from a vector score alone. Cross-model knowledge bases cannot be linked together; verify the identifier, exclude those that do not match, and record the actual observation rather than running a cross-model search to compare missed recall.