What the documents look like
Permit handling occurs before customs declaration and determines whether goods may be imported or exported lawfully. The materials in this step are mainly automatic import licences, dual-use item licences, contracts and product descriptions. They come from commerce authority systems, customs commodity codes, supplier materials and customer end-use statements. Each shipment requires fresh assembly, and updates are irregular rather than scheduled.
A single shipment may mix Chinese and English. Licence numbers, commodity codes and model designations are exact fields, while end-use statements and contract clauses are natural-language passages. Chunking and indexing must serve both kinds of content, because retrieval will otherwise fail to return the right passage reliably.
The documents also carry different evidence roles. A licence identifies a holder and a permitted scope. A contract records agreed terms. A product description explains what the item is and how it will be used. These roles affect how a chunk should be bounded: a condition and the content it constrains belong together, and a licence number should remain attached to the holder and validity period it qualifies.
Constraints at the document chunking step
FastGPT stores a dataset as a library, a collection and data. A group of data can map to more than one vector index, so retrieval can recall that group when any one index matches. This multi-vector mapping exists because content length and semantic richness pull against each other in one vector; splitting a long passage into several indexes preserves both coverage and precision.
Semantic retrieval suits natural-language questions and supports cross-language matching, for example a Chinese question matching an English knowledge point. Full-text retrieval suits numbers, models and proper nouns. Hybrid retrieval combines semantic and full-text recall and merges the two result sets through RRF. Rerank requires a model and is applied according to configuration. Minimum relevance is a filter threshold that takes effect only under semantic retrieval or result reranking; it does not isolate businesses, entities or unauthorized records. Retrieval is not authorization.
The library-to-vector-model association is limited to the knowledge base itself. It does not extend to external systems such as a commerce authority portal or a corporate registry. A library tied to one vector model cannot be jointly associated with a library using a different vector model; the identifier must be checked and non-matching libraries excluded, with the actual observation recorded rather than an inferred runtime result.
Image retrieval is an optional path, not a step in every request. It requires image input and matching model capability. Rerank likewise requires a model and must be enabled. Neither should be assumed present.
Configuration and implementation checks
| Configuration or check item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid retrieval | Materials contain both exact numbers and end-use prose; compare against semantic-only results |
| Reference limit | To be set from actual chunk length | The documented unit is tokens, not top k; compare against total candidate length |
| Minimum relevance | To be set; applies only under semantic retrieval or rerank | Filters low-relevance results; compare against rerank scores |
| Question optimization | Enabled as needed | Completes references in follow-up turns; compare against results with it off |
| Chunking review | Check licence, contract and product description separately | Keeps conditions with constrained content; compare against original document structure |
| Training status check | Verify actual training and index state after upload | Upload success does not equal updated retrieval; compare against retrieval observation |
Three common errors
- Treating a successful upload as an updated index. The check is to compare the training status and an actual retrieval observation, confirming the new chunk can be recalled.
- Treating minimum relevance as a permission or entity filter. The check is to compare the returned records against the query scope and confirm that this value only filters low-relevance results; it does not exclude unrelated businesses or unauthorized records.
- Associating libraries built on different vector models. The check is to compare library identifiers first, exclude non-matching libraries, and record the actual observation without inferring a runtime result.
How to verify the setup
Hypothetical paper test case. Source text for comparison: "Licence number: AUTO-2024-0117; importer: a machinery company in East China; commodity name: CNC grinding machine; licence valid until 31 December 2024; the licence holder is the machinery company in East China." A draft sentence awaiting judgement: "This licence may be used to export similar equipment to a trading company in South China." The comparison is between the draft sentence's subject, object and scope and the source text's stated holder and permitted scope. The source text does not mention the South China trading company or an export use, so the draft sentence is not supported.
- Input the licence number "AUTO-2024-0117" from the hypothetical sample into full-text or hybrid retrieval. Compare whether the returned content points to that data group. If no match appears, check training and index status. Observation and failure branch only; no guarantee of a first-hit match.
- Input the draft sentence from the hypothetical sample. Compare whether the results include the source text's stated holder and validity fields. If results contain only other companies' materials, record the outcome as unsupported. Do not make a permission determination.
- Use an explicit paper input of an empty result to check downstream design. Compare whether the downstream handles an empty reference as designed. Do not claim that a fictional rule will always return an empty array.
Each check records what was entered, what was compared and what was observed. Missing evidence remains missing. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination.