What the documents look like
Repair and reship work aims to restore a product to use, and it does not close through a direct refund. Product information and selection questions therefore have to point back to repair applications, reship orders, inspection reports and accessory outbound orders. These materials come from the after-sales system, the order system, repair outlets and the accessory warehouse WMS, and they update per order and daily.
Fast updates mean a citation has to state the document version and its valid scope. A retrieved passage is not automatically the currently valid passage. Selection questions about accessory models and compatibility relations involve numbers, model names and proper nouns, which suit full-text or hybrid retrieval, while natural-language fault descriptions suit semantic retrieval.
Constraints at the citation tracing step
The knowledge base search node can link one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. Even when nothing is found, the output path still runs through. An empty array is existing node behaviour, and it does not by itself show that the workflow errored or terminated. Retrieval results are handed to the answering step, where each answer is checked against the cited source text.
The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used, and it is a filter threshold, not a permission or entity isolation mechanism. A similarity score does not exclude unrelated businesses or unauthorized records. Hybrid retrieval merges multiple result paths through RRF, reranking depends on a model and runs when configured, and image retrieval needs image input plus the matching model capability. These paths do not all execute on every request.
Hypothetical paper exercise: a brand after-sales outlet holds a paper form reading "Repair application WX-2024-0311, product is Model A water purifier, fault described as odour in the outflow, customer requests a reship of the filter element, application date 11 March 2024, status awaiting outlet inspection." A second sheet reads "Accessory outbound order PK-2024-0312, accessory is Model A filter element, quantity 1 piece, outbound date 12 March 2024, purpose is reship." The draft sentence to be judged is "The repair application has approved the reship, the accessory has left the warehouse and customer confirmation is complete." The materials show only that the application status is awaiting inspection and that the accessory left the warehouse. They do not record approval, customer confirmation or installation completion.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Select the after-sales repair library and the accessory library | Both libraries must share the same vector model; check the identifier first and exclude any that do not match, without running a cross-model retrieval comparison |
| Search mode | Hybrid retrieval | Model names and order numbers need full-text hits while fault descriptions need semantic recall; the comparison object is a single semantic retrieval result |
| Citation limit | Hypothetical starting point of 2000 tokens | This value is a paper starting point for this page, used to compare against source text length; it is not an official default |
| Minimum relevance | Hypothetical starting point of 0.3 | Takes effect only in semantic retrieval or reranking; the comparison object is the citation count before and after filtering |
| Result reranking | Enabled as configured | Reranking requires a model; once enabled its score can be used for filtering, and a first-position hit is not assured |
| Citation display | Check each answer against the source text | Compare against the original text of the repair application, reship order, inspection report and accessory outbound order |
Three common errors
- Treating a retrieval hit as business authorization. The input is a repair application awaiting inspection plus an accessory outbound order, the comparison object is the phrase "has approved the reship", and the observable result is that the materials lack an approval field, so the draft sentence is not supported.
- Treating minimum relevance as a permission filter. The input is a similarity setting plus an unauthorized record, the comparison object is the returned retrieval content, and the observable result is that similarity does not exclude unauthorized records; authorization needs a downstream implementation.
- Treating an empty citation array as an error. The input is a clear paper input of [], the comparison object is the downstream handling design, and the observable result is that the node output path still runs through; whether an error occurs needs separate checking.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence "The repair application has approved the reship, the accessory has left the warehouse and customer confirmation is complete." Compare whether the source text contains an approval field or a customer confirmation field. The observable result is that neither is recorded, so the judgement is pending verification.
- Input accessory outbound order PK-2024-0312 and the question "Has the Model A filter element left the warehouse?" Compare whether the citation content contains that order number and outbound date. A hit that can be traced back satisfies the condition; a miss is recorded as a gap.
- Input a paper empty array []. Compare whether the downstream answering step accepts a zero-length citation. The path running through without displaying a source satisfies the condition; a downstream error is recorded as a design gap.
This page addresses only after-sales repairs and reships, product information and selection questions, and citation tracing. Whether materials, operations or requests are authorized at handover cannot be inferred from retrieval or from a form. The same vector model requirement for linked libraries does not extend to business registration sites or other external systems. RRF is a merge process for hybrid results and is not written as a toggle. Thinking budget is a reasoning budget and is not reused as a citation token limit. Optional paths in the overall flow diagram do not all run on every request: image retrieval needs matching image input and model capability, and reranking needs a model and runs when configured. Libraries with different vector models cannot be linked together; check the identifier, exclude those that do not match, and record what is observed. A judgement requires the same subject, the same period and the same field. Only mutually exclusive values in one field count as a contradiction. A field that is not recorded, or evidence that shows only a related stage, is unsupported or pending verification, not a negation or a conflict. Pending review does not show a lack of operating permission; signing, reaching cooperation and contract effectiveness are three separate facts; a quotation validity period does not state a supply period.