What the documents look like
Permit handling for import and export takes place before declaration and determines whether goods can legally cross the border. Operating reports and exception reviews therefore depend on fixed citation sources for automatic import licences, dual-use item licences, contracts and product descriptions, so that each shipment can be traced back. Data sources named in the brief include commerce authority systems, customs commodity codes, supplier materials and customer end-use statements; updates are irregular and per shipment. These materials sit in different systems, and every conclusion in a report needs to point back to a specific document and field.
For citation tracing, the practical consequence is that a retrieved passage is only a candidate source. Whether it is adopted still depends on a valid-version review. Retrieval does not perform authorization checks, and the minimum relevance value is a filtering threshold rather than a category, entity or permission filter. Mixed retrieval combines multiple result paths through RRF, reranking requires a model and must be enabled in settings, and image retrieval requires image input plus the corresponding model capability. Knowledge bases with different vector models cannot be linked together; identifiers are checked first and non-matching bases are excluded.
Constraints at the citation tracing step
The knowledge base search node accepts one or more knowledge bases that share the same vector model and outputs cited content as an array, which may have a length of zero. A zero-length output means the downstream chain still runs; it does not mean the workflow has failed or terminated. Checking whether a retrieved passage was adopted remains a separate step from retrieval itself.
A hypothetical paper exercise sample is used below. It is clearly labelled and is not a real customer record.
Sample source sentences:
- "Automatic import licence application number A-2024-0517, applicant an electromechanical import enterprise in East China, goods CNC machine tool accessories, declaration status pending review."
- "Dual-use item licence application number B-2024-0088, same applicant, goods high-precision sensors, signature status unsigned."
- "Purchase contract HT-2024-0312 provides for delivery in two batches; the first batch delivery date is 30 June 2024; the quotation validity period is 1 May 2024 to 31 May 2024."
Draft sentence to be assessed: "Numbers A-2024-0517 and B-2024-0088 have both been approved, and contract HT-2024-0312 has a supply deadline of 30 June 2024."
The draft sentence conflicts with the source on status fields for A-2024-0517 and B-2024-0088, and it treats the quotation validity period as if it were a supply deadline. The source does not state approval, signature or a supply deadline, so those parts remain unsupported rather than disproved.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Only permit material bases sharing one vector model | Compared with the identifier of a base using a different vector model; non-matching bases are excluded |
| Search mode | Mixed retrieval | Compared with semantic retrieval alone; number-type queries are checked for full-text hits |
| Citation upper limit | Hypothetical starting value 3000 tokens | Compared with the displayed default length; adjusted by the length of each shipment's materials |
| Minimum relevance | Hypothetical starting value 0.5 | Compared with the retrieval score distribution; effective only under semantic retrieval or reranking |
| Result reranking | Enabled | Compared with the disabled state; secondary ordering is checked when candidates are numerous |
| Citation display check | Compare each answer line against the cited source text | Compared with the thinking process; the source text is the evidence |
The values in the table are hypothetical starting points for this exercise, not official defaults. Thinking budget and citation token limit are two different parameters and are not interchangeable.
Three common errors
- Treating "pending review" as approved. Input the draft sentence together with the source sentence for A-2024-0517, compare the declaration status field, and observe that the source states "pending review" while the draft asserts "approved". The draft's approval claim is unsupported by the source.
- Treating "unsigned" as effective. Input the source sentence for B-2024-0088, compare the signature status field, and observe that "unsigned" does not constitute a completed signature or entry into effect. The draft's approval claim is unsupported.
- Treating the quotation validity period as a supply deadline. Input contract HT-2024-0312, compare the quotation validity period field with the delivery date field, and observe that they are distinct fields. The draft's "supply deadline of 30 June 2024" is unsupported by the source.
How to verify the setup
- Input the three source sentences above and the draft sentence, compare subject, number and status field by field, and observe whether the output identifies the inconsistency between "approved" in the draft and "pending review" in the source. Expected result: the output points to the specific field mismatch rather than restating the draft.
- Input a clearly written empty array [] on paper, and check whether the downstream answer stage continues with empty citations. Expected result: the downstream stage runs through with zero-length cited content; this check does not claim that a fictional rule search will always return [].
- Input two knowledge base identifiers that share one vector model and one identifier that uses a different vector model, compare the identifiers, and observe whether the non-matching base is excluded and the actual observation is recorded. Expected result: only same-model bases remain linked; no cross-model retrieval is executed to compare missed recall.
Across all three checks, retrieval and forms collect or return values; business verification, approval, database writes and external synchronization require their stated downstream implementation. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where a document field is not provided, the value remains unknown rather than assumed.