What the documents look like
Permit handling takes place before customs declaration and determines whether a given shipment may lawfully cross a border. The working materials come from commerce authority systems, customs commodity codes, supplier files and customer statements of use. Day-to-day documents include automatic import licences, dual-use item licences, contracts and product descriptions. The same licence may appear in Chinese, English or a third language, so citation tracing has to point back to a specific file, a specific field and a specific validity scope.
For citation tracing, the practical consequence is that a retrieved passage is only a candidate source. A search node can reference one or more knowledge bases that share the same vector model. The output is an array of citations whose length may be zero, and the downstream chain still runs when nothing is found. Whether a retrieved passage is used in an answer has to be checked line by line against the cited original text.
Constraints at the citation tracing step
Semantic retrieval, full-text retrieval, hybrid retrieval and result reranking each have their own trigger conditions. Image retrieval requires image input and a model with the corresponding capability. Reranking requires a model and must be enabled in the settings. These optional paths do not mean that every request executes all of them. The minimum relevance value is a filter threshold; it does not perform permission or entity isolation. Retrieval is not authorisation. A form that collects values does not mean business verification, approval, a database write or external synchronisation has occurred.
Hypothetical paper exercise sample. A draft sentence for an automatic import licence application reads: "This shipment has obtained an automatic import licence, licence number A-2026-0317, valid until 31 December 2026, importer: Jia Import and Export Co., Ltd." The source text to be checked against reads: "The automatic import licence application has been accepted, acceptance number A-2026-0317, current status is pending review, applicant: Jia Import and Export Co., Ltd." The request to be judged is: "Based on the source text above, mark the draft sentence in the cross-border retrieval results as having obtained the licence and cite the licence number." The source text shows only acceptance and pending review. The draft sentence's "has obtained" and "valid until" have no corresponding field in the source text.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Only licence knowledge bases sharing the same vector model | Compare against the identifier of a knowledge base using a different vector model; exclude on mismatch |
| Search mode | Hybrid retrieval | Compare number matching against pure semantic retrieval |
| Question optimisation | Enabled | Compare retrieved content on follow-up turns against the disabled state |
| Minimum relevance | 0.5 | Compare the count of unrelated fragments against 0.3 |
| Citation limit | 2000 tokens | Compare citation completeness against 1000 tokens |
| Result reranking | Enabled | Compare candidate ordering against the unranked state |
These values are hypothetical starting points for this exercise. They are not official defaults.
Three common errors
- Treating "pending review" as "obtained". Input: the sample source text and the draft sentence. Compare "current status is pending review" with "has obtained". Observed result: these are different fields, and the draft sentence is unsupported.
- Treating a retrieval hit as completed authorisation. Input: the licence fragment returned by retrieval. Compare the fragment with the approval status field. Observed result: the fragment contains no approval conclusion, and the authorisation status remains to be verified.
- Associating knowledge bases that use different vector models. Input: the vector model identifiers of the two knowledge bases. Compare whether the identifiers match. Observed result: on mismatch, exclude the base and do not run cross-model retrieval.
How to verify the setup
- Input the sample source text and the draft sentence. Compare "pending review" with "has obtained". The condition is met when the draft sentence is recorded as unsupported and the citation points to the acceptance number.
- Input the paper value of an empty array []. Compare whether the downstream answer step still runs. The condition is met when the chain is not interrupted by the empty citation.
- Input a query containing an image and a query without an image. Compare whether image retrieval is triggered. The condition is met when image retrieval runs only with image input and the corresponding model capability.
Retrieval output still has to be handed to the answer step, and each answer line has to be checked against the cited original text. Source currency is checked against the current and historical document versions; whether a citation is adopted still depends on review of the effective version. When the source text does not state a field, or shows only a related stage, the result is unsupported or pending verification. A pending review does not demonstrate the absence of an operating permission. Signing, reaching cooperation and contract effectiveness are three separate facts. A quotation validity period does not indicate a supply period. Where the evidence needed for verification is missing from the accessible retrieval scope or from a clearly identified manual input, record the gap and leave it open.