What the documents look like
Import and export permit handling takes place before customs declaration, and it decides whether a given shipment may lawfully cross the border. That timing shapes the document set. The working materials are automatic import licences, dual-use item licences, contracts, and product descriptions. They are read together with supplier records and customer end-use statements so that licence face fields can be checked against the goods actually being purchased.
In stock procurement and supplier files, one supplier may hold several contracts and several product descriptions at once. The link between a licence face field and a contract goods description is spread across separate files. The retrieval unit that matters is therefore "the permit status of this shipment from this supplier," treated as one object. In FastGPT the smallest search unit is the library; a collection is only a classification layer and does not affect search results.
Data sources named for this category include the commerce authority system, customs commodity codes, supplier materials, and customer end-use statements. Updates occur irregularly, per shipment. This page does not treat that cadence as a trigger for batch or real-time synchronisation.
Constraints at the retrieval step
When associating knowledge bases, only one or more libraries that use the same vector model can be selected together. Libraries built on different vector models cannot be associated jointly. The practical step is to check the vector model identifier first, exclude any library that does not match, and record what was actually observed. No conclusion about recall behaviour after a cross-model association should be drawn, because that association is not available.
Semantic retrieval compares vector distance and suits natural-language descriptions. Full-text retrieval relies on keyword hits and suits licence numbers, commodity codes, and model designations. Hybrid retrieval uses both paths and merges them through RRF. RRF is a fusion process inside hybrid retrieval, not a switch that an operator turns on. Result re-ranking requires a re-rank model and must be enabled in settings before ordering can be compared. Image retrieval requires image input and matching model capability. These are optional paths listed in the documentation; their presence does not mean every request executes them.
The minimum relevance value is a filter threshold and applies only in semantic retrieval or when result re-ranking is used. It does not perform permission or entity isolation, and retrieval itself is not authorisation. Reference content is output as an array whose length may be zero; when nothing is retrieved, the output path still completes.
Hypothetical paper exercise sample
The following sample is a paper exercise for checking method only. It does not refer to any real party.
Source text (hypothetical): "Supplier: East China Mechanical Fittings Works (hypothetical name). Contract number HX-2024-031, signed 12 March 2024, goods are a hydraulic valve body, quantity 200 pieces. An automatic import licence is attached; the commodity code on the licence face matches the contract goods description; the licence is valid until 30 September 2024."
Draft sentence to be assessed (hypothetical): "This supplier has obtained the automatic import licence for this shipment and shipment arrangements may proceed."
Assessment point: the draft joins "licence obtained" with "shipment may proceed." The source text supports the existence of the licence and its validity range. It does not state shipping authorisation, completed approval, or a release conclusion. Under the same party, same period, and same field, licence validity and shipping arrangements are different facts. The second half of the draft sentence is unsupported and should remain pending verification.
Configuration and implementation checks
| Configuration item | Value or check for this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Select libraries with the same vector model identifier; check each before associating | Same vector model is the precondition for joint association; exclude non-matching libraries and record the observation |
| Search mode | Compare semantic, full-text, and hybrid results separately against the test question | No single mode is assumed to suit all materials |
| Reference limit | Starting point for this hypothetical trial is 2000 tokens, compared with 1000 and 4000 for content completeness | Unit is token; the starting point is a trial contrast, not an official default |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold step by step | Applies only in semantic retrieval or with result re-ranking |
| Question optimisation | Compare the original question with the optimised question and check whether the subject was altered | Used for coreference resolution and question expansion; adds one model call |
| Result re-ranking | Confirm a re-rank model is available and enabled in settings before comparing order | Re-rank scores can be used for filtering; this is an optional path |
Three common errors
- Treating a retrieval hit as authorisation. Input is a query about one shipment from one supplier; the comparison object is the retrieved content against the authorisation status field; the observable result is licence information appearing in the reference content. Concluding that shipment may proceed turns a retrieval result into an executed authorisation. The correct action is to record the retrieval hit and verify shipping authorisation separately.
- Forcing joint association of libraries on different vector models. Input is two libraries with different vector model identifiers; the comparison object is the model identifier; the observable result is that the identifiers do not match. The non-matching library should be excluded and the observation recorded. Descriptions of missed recall or redundant results after such an association should not be produced.
- Writing a pending state as completed. Input is the draft sentence "licence obtained, shipment arrangements may proceed"; the comparison object is the validity range stated in the source against the shipping authorisation not stated; the observable result is that the source lacks any shipping authorisation wording. The second half of the sentence should be marked pending verification and should not be described as a denial or a conflict.
How to verify the setup
- Verification one: use the hypothetical source text and draft sentence as input, and compare the draft phrase "shipment arrangements may proceed" with the licence validity field stated in the source. The condition is met when the assessment states that the licence existence is supported by the source, that shipping authorisation is unsupported and remains pending verification, and that no "shipment approved" conclusion appears.
- Verification two: use two libraries with different vector model identifiers as input, and compare the model identifier field. The condition is met when the check record shows the identifiers do not match, the non-matching library is excluded, and no description of missed recall from a cross-model retrieval comparison appears.
- Verification three: use an empty reference array as input, and check how the downstream design handles reference content of length zero. The condition is met when the output path completes, the downstream treats it as an empty result, and no claim is made that a fictional rule causes retrieval to return an empty array.
The verifications above are paper expectations. They were not executed and do not constitute a test, deployment, or approval conclusion.