What the documents look like
Permit handling takes place before customs declaration and determines whether a shipment may lawfully cross the border. The materials in this category are auto-import licences, dual-use item licences, contracts and product descriptions. They arrive from commerce authority systems, customs commodity codes, supplier files and customer end-use statements. The comparison task requires checking the goods, quantity, end use and validity period stated on a licence against contract clauses and product descriptions. These three material types come from different parties at different times, and their field names and granularity do not match. Retrieval therefore gathers relevant passages for human review. Retrieval does not produce the comparison conclusion.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. Check the vector model identifier first, exclude non-matching bases, and record what was observed. Reference content is output as an array and its length may be 0. When nothing is retrieved, the output path still runs.
Search mode is chosen by question type. Exact queries such as numbers, models and licence numbers suit full-text search or hybrid search. Natural-language descriptions suit semantic search or hybrid search. Hybrid search merges result paths through RRF, which is a fusion process rather than a switch that is turned on. Result reranking requires a ReRank model and must be enabled in settings; the rerank score is 0-1. Minimum relevance takes effect only in semantic search or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation, and it does not act as the final ordering rule. The reference cap is measured in tokens and limits how much content a single search may cite.
Question optimization performs coreference resolution and question expansion, adding one model call before the formal retrieval. Whether to enable it depends on observed results.
Hypothetical paper exercise sample: a shipment is proposed for import. The supplier's contract states "Commodity name: ball screw for CNC machine tools, quantity 200 sets, delivery date March 2025." The product description states "End use: transmission component for CNC machine tools." A draft sentence for the auto-import licence reads "This certifies that the above goods have obtained an auto-import licence, valid until June 2025." The request to be judged is whether this draft sentence can serve as handover evidence that the licence has been obtained. The licence original is not present in these materials, and neither issuance status nor licence number is stated.
Configuration and implementation checks
| Configuration item | Value or check for this page | Basis and notes |
|---|---|---|
| Associated knowledge bases | Select the permit materials base and the contract materials base; first check that their vector model identifiers match | The node can associate one or more knowledge bases using the same vector model |
| Search mode | Use full-text or hybrid search for field queries such as "ball screw quantity 200 sets" | Numbers, models and proper nouns suit full-text or hybrid search |
| Reference cap | For this hypothetical trial, start at 2000 tokens and compare against 1000 tokens for whether content stays complete | Unit is tokens; adjust against actual material length |
| Minimum relevance | For this hypothetical trial, start at 0.3 and compare against 0.5 for whether relevant passages are retained | Takes effect only in semantic search or result reranking |
| Question optimization | Enable for referring expressions such as "the above goods"; compare whether the subject is altered before and after | Used for coreference resolution and question expansion |
| Result reranking | Enable ReRank when candidates are numerous; filter by the 0-1 score | Requires a rerank model and must be enabled in settings |
Three common errors
- Treating a retrieval hit as proof that a licence has been obtained. In the sample, the draft sentence claims an auto-import licence has been obtained, while the materials contain only a contract and a product description. Licence original, issuance status and licence number are all unstated. A retrieved contract passage is relevant-stage evidence and does not support the issuance fact.
- Treating minimum relevance as a permission or entity filter. A setting of 0.3 only affects which semantic or reranked results are retained. It cannot exclude unrelated businesses or unauthorized records. Retrieval does not perform authorization judgment.
- Forcing knowledge bases on different vector models into one association. When two bases have non-matching vector model identifiers, exclude the non-matching one and record the observation. Do not run cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the contract text from the sample, "Commodity name: ball screw for CNC machine tools, quantity 200 sets, delivery date March 2025," and compare whether the returned passages contain the commodity name and quantity fields. If the returned content contains those fields and remains complete within the reference cap, the condition is met. If the return is empty or a field is truncated, record it as not met.
- Input the draft sentence "This certifies that the above goods have obtained an auto-import licence, valid until June 2025," and compare whether the returned content contains licence issuance status or licence number. If the returned content contains neither issuance status nor licence number, judge it as unsupported and preserve the evidence gap. If the returned content contains an issuance record, check whether it concerns the same party, the same period and the same field.
- Input the explicit paper empty result
[]and check whether the downstream design continues along the empty-reference path. If the output path runs and downstream does not treat the empty result as verified, the condition is met. If downstream treats the empty result as an obtained licence, record it as a design defect.