What the documents look like
Brand after-sales repair work is organized around warranty responsibility and physical service locations. The records that matter here are warranty cards, repair work orders, inspection reports, and replacement approvals. In a cross-border, multilingual setting, one repair work order can carry a Chinese fault description, an English spare-part number, and a local-language note from a service location. Retrieval therefore has to serve two kinds of questions at once: natural-language descriptions and exact identifiers such as part numbers, model codes, and report numbers.
These records are updated per order and daily. That update rhythm does not by itself create any batch, real-time, or transaction-triggered mechanism in the retrieval step. What matters for this page is the field scope of a single work order and whether a search can reach the right record inside the connected knowledge base.
Constraints at the retrieval step
A knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. The practical action is to check the vector model identifier and exclude any knowledge base that does not match, then record what was actually observed. Running a cross-model search to compare missed recalls is not part of this step.
Semantic search suits natural-language descriptions. Full-text search suits numbers, model codes, and proper nouns. Hybrid search combines the two paths through RRF. The minimum relevance value takes effect only in semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion, and it adds an extra model call before retrieval.
The search node outputs cited content as an array, and that array can have a length of zero. When no content is found, the output path still runs. This is a documented behavior of the node, not a guarantee about any particular query.
Hypothetical paper exercise: a brand after-sales center receives a cross-border repair work order. The source text contains three sentences: "Device model AX-220; warranty card records a purchase date of 10 March 2024 and a warranty period of 12 months." "Fault description: no display after power-on; inspection completed by the local service location, inspection report number R-0917." "Replacement approval status is pending review; the applicant has not yet signed." The draft sentence to be assessed is: "This work order has passed replacement approval and a replacement can be arranged." The exercise compares the draft sentence against the source text on the approval status field.
Configuration and implementation checks
| Configuration item | Starting value for this hypothetical exercise | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Link one after-sales record knowledge base that uses the same vector model | Check the vector model identifier first and exclude knowledge bases on a different vector model |
| Search mode | Try hybrid search first | Compare against semantic search alone and full-text search alone on identifier questions and natural-language questions |
| Citation limit | Start at 2000 tokens | Test with short and long work order texts to see whether cited content is complete, and compare against a shorter limit |
| Minimum relevance | Start at 0.3 | Label relevant and irrelevant samples first, then compare filtering results against higher and lower thresholds |
| Question optimization | Off at first | Compare the original question with the optimized question and check whether the subject was changed incorrectly |
| Result reranking | Not enabled at first | Confirm a rerank model is available and enabled by setting, then compare ordering against results without reranking |
The values in this table are starting points for this hypothetical exercise. They are not official default values. The citation limit is measured in tokens and is unrelated to the top-level thinking budget.
Three common errors
- Treating a semantic recall score or relevance score as a permission or entity filter. Retrieval does not perform authorization. The minimum relevance value is a filtering threshold and does not exclude unrelated business records or unauthorized records. Check: submit a query that spans service locations, compare whether the returned content still includes records from other locations, and observe whether an unauthorized citation appears.
- Treating pending review as approved. In the exercise source text, "Replacement approval status is pending review; the applicant has not yet signed" and the draft sentence "has passed replacement approval" are mutually exclusive values on the same field. Check: compare the draft sentence against the source text word by word on the approval status field, and observe whether the value "passed" appears where the source text does not state it.
- Describing RRF as a switch that can be turned on. RRF is the fusion process for hybrid search results and is not a separate configuration item. Check: inspect the configuration interface for a switch named RRF, and observe whether the fusion process has been mislabeled as a toggleable option.
How to verify the setup
- Input the exercise source text and the draft sentence, compare the draft sentence "has passed replacement approval" against the source text "pending review; the applicant has not yet signed," and observe whether the draft is judged unsupported. Keep the evidence gap open.
- Input a full-text search question containing "AX-220," compare the full-text search return against the semantic search return, and observe whether the identifier is matched exactly. Do not expect a first-position hit or a complete recall.
- Input an explicit paper-level empty result
[], check whether the downstream path still runs, and observe whether the output path can still carry the flow. Do not claim that a fictional rule search necessarily returns[].
The verification questions above test the named primary task of retrieval for multilingual work: whether a draft sentence is supported by the source record, whether an exact identifier is matched by the chosen search mode, and whether an empty citation array still allows the downstream path to run. Each check states its input, its comparison, and the observable result. None of them establishes that a test, deployment, approval, refund, or other business action has occurred.