What the documents look like
After-sales repairs and reships aim to restore product use. A case closes when the repair is completed or the replacement part reaches the customer. Direct refunds sit outside this page's primary task. The working materials come from the after-sales system, the order system, repair sites, and the parts warehouse WMS. Day-to-day documents include repair requests, reship orders, inspection reports, and parts outbound orders. These documents update per order and daily, so any sentence drafted for customer-facing copy must match the current state of the document it draws on.
The document shapes differ in ways that matter at retrieval. Repair requests and inspection reports are mostly natural-language narratives, so semantic retrieval can cover paraphrases. Reship orders and parts outbound orders carry part numbers, model designations, and document numbers, so full-text retrieval depends on keyword hits. A draft sentence about repair scope, reshipped parts, or timing can only rest on the inspection report and reship order text already stored in the knowledge base. The retrieval step recalls the relevant material. It does not decide whether that material has been approved.
A second constraint comes from the retrieval mechanism itself. The knowledge base search node can associate one or more knowledge bases that use the same vector model. A knowledge base built on a different vector model cannot be associated alongside them. The reference upper limit is measured in tokens, and mixed long and short materials make it worth testing content completeness with actual documents. The minimum relevance value applies only under semantic retrieval or when result reranking is used, and it acts as a filter threshold. It does not isolate unauthorized records or unrelated businesses. When no content is found, the reference output is still an array and its length can be 0, and the downstream link still runs.
Constraints at the retrieval step
The constraints above shape how a draft sentence is checked against source material. Semantic retrieval, full-text retrieval, and hybrid retrieval behave differently on the same question, and no single mode suits every document type. Hybrid retrieval merges two recall paths through RRF. Reranking requires a rerank model and runs according to its settings. Question optimization performs coreference resolution and question expansion before retrieval, and it adds one model call, so it usually improves multi-turn retrieval while increasing overall latency.
A hypothetical paper exercise sample illustrates the check. It is clearly labelled and is not a real customer record.
Sample material text, stated as a hypothetical exercise: "Reship order BS-2026-0417 states: the left pedal part for customer order A-88231 left the parts warehouse on 17 April 2026, quantity 1 piece. Inspection report JC-2026-0392 states: the pedal fracture resulted from transport impact, not human damage, and reship is recommended. Repair request WX-2026-0511 states: the customer submitted the reship request on 15 April 2026, and the current status is pending site confirmation."
Draft sentence to be judged: "The parts for this order have been confirmed for reship and are expected to arrive within three days."
The comparison target: whether "confirmed for reship" in the draft sentence matches the current status "pending site confirmation" in the material. The material records a pending state. The draft sentence asserts a confirmed state. The two values are not the same field with mutually exclusive values; the material simply does not carry a confirmation. The draft sentence is unsupported by the material, and the gap stays open.
Configuration and implementation checks
The configuration below stays tied to this sample and to the retrieval feature and copy-drafting task. The values are hypothetical starting points for this exercise, chosen for comparison, and are not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One after-sales repair and reship knowledge base | The sample material belongs to this base; confirm the base uses the same vector model before associating it |
| Search mode | Hybrid retrieval | The draft sentence contains natural-language wording while the material contains document numbers and part names; compare semantic and full-text recall paths |
| Reference upper limit | 1200 tokens | The three sample sentences total a few hundred characters; use this value to observe whether the document number and status sentence enter the reference completely |
| Minimum relevance | 0.35 | Annotate relevant and irrelevant sentences in the sample first, then compare what remains after filtering |
| Question optimization | Enabled | The draft sentence contains the reference "this order"; compare whether the optimized question still points to reship order BS-2026-0417 |
| Result reranking | Enabled according to settings | When candidates are numerous, compare rerank scores and reference order, and confirm the status sentence is not pushed outside the reference |
Three common errors
- Treating "pending site confirmation" in the material as confirmed reship. The erroneous input is the draft sentence "the parts for this order have been confirmed for reship." The comparison object is the current status field in the reship order and the repair request. The observable result is that the two status values do not match, and the draft sentence is unsupported by the material.
- Using the minimum relevance value to exclude unrelated businesses or unauthorized records. The erroneous input treats the relevance threshold as a permission condition. The comparison object is the documented scope of the threshold against the access-control implementation. The observable result is that the document states this value filters low-relevance results only, and permission requires a downstream implementation.
- Associating knowledge bases that use different vector models and then inferring missed recall. The erroneous input is two knowledge bases built on different vector models. The comparison object is the same-model requirement in the association condition. The observable result is that the identifiers should be checked first and non-conforming bases excluded, and no runtime result is asserted.
How to verify the setup
- Input the three sample source sentences and the draft sentence. Compare "confirmed for reship" in the draft sentence against "pending site confirmation" in repair request WX-2026-0511, and observe whether the reference content includes that status sentence. If the reference contains only the reship order and not the status sentence, the draft sentence is unsupported; raise the reference upper limit or adjust the search mode and retry.
- Input the draft sentence with question optimization enabled. Compare whether the optimized question still points to reship order BS-2026-0417. If the optimized subject changes to another document number, record the observation and check the question optimization setting. Do not conclude from this that retrieval failed.
- Input a clearly paper-based empty-result scenario. Compare whether the reference content array length is 0 and whether the downstream link still runs. If the downstream does not accept an empty array by design, record this as a downstream implementation gap. Do not assert that the retrieval rule returns an empty array.
These checks are paper expectations. They were not executed and do not constitute a test, deployment, or approval conclusion.