What the documents look like
Brand after-sales repair work is organized around warranty responsibility and physical service points, so the material set differs from a platform refund desk. The documents handled here are warranty cards, repair work orders, inspection reports and exchange approvals. They arrive from the after-sales system, store orders and service point feedback, and they change per order and daily. Warranty scope decisions, spare-part coordination and exchange or return approval consume the most staff time, and a poor retrieval result shows up as a consumer complaint, an overrun repair or a compensation case.
For this page the retrieval step reads the documents attached to the current request. It does not perform batch movement, scheduled synchronization or transaction-triggered updates, and the daily or per-order update rate of the source material does not by itself add any such mechanism here.
The knowledge base search node can attach one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be attached together; the model identifier has to be checked and the non-matching base excluded before any search is run. Semantic retrieval, full-text retrieval and hybrid retrieval serve different question shapes. Hybrid retrieval merges the routes through RRF, ReRank performs a second ordering of candidates, the citation cap is measured in tokens, and the minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference removal and question expansion and adds one model call. Retrieval and relevance scores do not carry permission or entity isolation, and an empty reference list still lets the downstream chain run.
A hypothetical paper exercise for this page: a brand after-sales center holds a paper warranty card reading "Warranty card A-001, product model X200, purchase date March 2024, whole-unit warranty period 12 months, mainboard within warranty scope, cosmetic wear outside warranty scope." The draft sentence under review is "The A-001 mainboard fault applied for repair in April 2025 and falls within warranty scope." The application date in that sentence sits outside the period stated on the card, which ends in March 2025, so the sentence is not supported by the source text.
Constraints at the retrieval step
The same-vector-model requirement belongs to the knowledge base attachment field. It does not extend to any external system, and it does not mean that an attached base will return every matching record. The citation cap is a token budget for the content returned in one search, not a count of items, and it is not the same thing as a thinking budget used by a model.
The minimum relevance value is a filter threshold, not a final ordering rule. After question optimization, hybrid retrieval, image retrieval or result reranking, the final set may come from several merged routes and will not necessarily follow a single vector similarity order. A relevance score is not a category filter, an entity filter or a permission filter, and it does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization.
Forms in this business collect values. Business verification, approval, database writes and external synchronization each need their own downstream implementation, and none of them is created by a retrieval setting. A pending approval is not an approval. An offered option is not execution. An unsigned termination is not a completed termination.
The full retrieval diagram lists optional paths, and not every request exercises all of them. Image retrieval requires an image input and the relevant model capability. Reranking requires a rerank model and its configured activation. Where a document field is not supplied, its value is unknown, and no default, limit, unit, menu or global switch is assumed here.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Link to the exercise sample |
|---|---|---|
| Attached knowledge bases | Attach one or more knowledge bases sharing the same vector model | The warranty card and the work order must sit in bases with the same vector model |
| Search mode | Compare semantic, full-text and hybrid retrieval against the question | An exact identifier such as model X200 suits full-text or hybrid |
| Citation cap | Measured in tokens; test completeness with short and long material | The warranty card is short and the work order is long, so each is checked separately |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Coreference removal and question expansion, adding one model call | In a follow-up such as "that unit", check whether the subject was altered |
| Result reranking | Present when a ReRank model exists and is activated by configuration | With many candidates, the second ordering yields a 0-1 score usable for filtering |
For this page the hypothetical trial starting value for the citation cap is 2000 tokens. The reason is that the warranty card and the work order together are short, and the comparison is against a 1000-token starting value to see whether content is truncated. This value is a trial starting point and is not an official default.
Three common errors
- Attaching knowledge bases that use different vector models. The input is two bases whose model identifier fields differ. The comparison is between those identifier fields. When they do not match, the base is excluded, and no cross-model search is run to compare missed recalls.
- Treating the minimum relevance value as a permission or entity filter. The input is a question about an unrelated business. The comparison is whether the returned content belongs to the authorized target scope. A relevance score does not exclude unauthorized records, and retrieval itself is not authorization.
- Treating a pending exchange approval as a completed exchange. The input is a work order whose approval status reads pending. The comparison is between the status field and the completion field. A pending status does not show that an exchange was carried out, and it does not show that operating permission is absent.
How to verify the setup
- Input the hypothetical sample text above together with the draft sentence "The A-001 mainboard fault applied for repair in April 2025 and falls within warranty scope." Compare the application date with the warranty end date. When the application date is later than the end date, the sentence is not supported, and the expected behavior is that the draft sentence is not output as a within-warranty conclusion.
- Input the model identifier "X200" through full-text or hybrid retrieval. Compare whether the returned content contains that model field. When no candidate carries the exact identifier, record the failure branch. A first-position hit is not guaranteed.
- Input an empty paper input [] and compare whether the downstream chain still runs. The expected reference content length is 0 and the output chain continues. This does not claim that a search under a fictional rule necessarily returns [].
Across all three checks, missing evidence stays missing. The verification covers the named primary task of retrieval for product questions against the supplied source wording and its stated limits, and it does not stand in for identity certification, report analysis or any other task.