What the documents look like
Brand after-sales repair work is organised around warranty responsibility and physical repair outlets, so the material handled goes beyond platform refunds. After-sales systems, store orders, repair outlet feedback and the national three-guarantees catalogue form the main data sources. Day-to-day documents include warranty cards, repair work orders, inspection reports and replacement approvals. Inventory procurement and supplier records are one business record group inside that set, covering spare-part names, supplier names, purchase requests, arrival records and authorization status. These records update per order and daily.
For retrieval, two features of this material matter most. The same spare part can be written differently across work orders, so a single wording may miss records that use a variant. Supplier records and purchase records are also stored separately, so a question about one supplier may need evidence from more than one library. Warranty scope decisions, spare-part coordination and return or exchange approvals are the most time-consuming steps, and irrelevant material pulled into the cited content adds manual cross-checking later.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Libraries built on different vector models cannot be linked together. The vector model identifier must be checked first, non-matching libraries excluded, and the actual observation recorded. Cross-model retrieval must not be run to compare missed recalls.
Search mode can be semantic retrieval, full-text retrieval or hybrid retrieval. No single mode suits all record types. Full-text retrieval suits precise queries such as numbers, models and proper nouns. Semantic retrieval suits natural-language questions. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process, not a standalone switch.
The citation limit is measured in tokens and caps how much content one search may cite. Minimum relevance is a filter threshold that takes effect only under semantic retrieval or when result reranking is used. It does not perform permission or entity isolation. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Cited content is returned as an array and its length can be zero, so the output path still completes when nothing is found.
Hypothetical paper exercise sample. A supplier record reads: "Supplier A, supply scope: compressor spare parts, authorization status is pending review, quotation valid until 30 June 2025." A purchase request reads: "Request to purchase 20 compressor spare parts, request date 10 May 2025, requester: store service agent B." A draft sentence awaiting judgment reads: "Supplier A is authorized to supply compressor spare parts." Comparison of the draft against the source shows the authorization status field holds a different value in each: the source says pending review, the draft says authorized. A quotation validity date is not a supply period, and pending review is not authorization.
Configuration and implementation checks
The table lists documented settings or implementation check items tied to this feature and this task. The values are starting points for this hypothetical exercise only; each needs a stated reason and a comparison object.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Supplier records library, purchase request library | Both must use the same vector model; check the vector model identifier first, exclude non-matching libraries, and do not run cross-model retrieval comparisons |
| Search mode | Hybrid retrieval | The sample contains a proper noun such as "compressor spare parts" and natural-language phrasing; compare candidate content under semantic, full-text and hybrid modes |
| Citation limit | 800 tokens | Referenced against the sample source length and typical work order length; compare whether cited content stays complete at different limits, then adjust |
| Minimum relevance | 0.35 | Label relevant and irrelevant fragments in the sample first; compare whether irrelevant fragments are filtered at different thresholds; effective only under semantic retrieval or reranking |
| Question optimization | Enabled | The sample is a single clear question; compare the original question with the optimized question, check whether the subject was altered, and observe the change in elapsed time |
| Result reranking | Enabled per settings | A rerank model must be available; compare candidate ordering before and after reranking; the rerank score is used for filtering, not for permission decisions |
Three common errors
- Treating pending review as authorized. The sample source says "authorization status is pending review." A draft sentence that says "is authorized" holds a different value in the same field, which is an error. Retrieval and forms collect values; they do not execute authorization confirmation. Authorization status must be checked against the supplier record source.
- Treating quotation validity as a supply period. The sample source says "quotation valid until 30 June 2025." That field describes the validity range of the quotation and does not state a supply period or contract effectiveness. A draft that builds a supply arrangement on it goes beyond what the source supports.
- Linking knowledge bases that use different vector models. The sample involves two libraries. If their vector model identifiers differ, they cannot be linked together. Check the identifiers, exclude non-matching libraries, and record the observation. Do not invent runtime outcomes of a joint link.
How to verify the setup
- Input the sample source and the draft sentence "Supplier A is authorized to supply compressor spare parts." Compare the draft against the source authorization status field. The source says pending review and the draft says authorized, so the same field holds different values and the draft is unsupported. The condition is met only if the source authorization status reads approved and the draft matches it.
- Input the sample source and the request "Please arrange compressor spare-part supply according to the quotation validity date." Compare the field the request relies on against the source. The source states only a quotation validity date and no supply period, so this is an evidence gap. The condition is met only if the source separately states a supply period field and the request matches that field.
- Input a paper empty result [] and check whether the downstream design handles a cited-content array whose length can be zero. Do not claim that a fictional rule makes retrieval always return []. If the downstream design explicitly accepts an empty array, the condition is met; if the downstream design depends on a non-empty result, record the gap. Retrieval checks record observations and failure branches only, and do not guarantee a first hit, full recall, or an empty result for an unrelated query.
At operational handover, retrieval and form collection of values do not constitute authorization execution. Pending review does not prove that operating permission is absent. Signing, reaching cooperation and contract effectiveness are three distinct facts. Fields that are unstated or supported only by a related stage remain unsupported or pending verification.