What the documents look like
Producer export tax processes combine the offset of domestic sales tax with the refund of export tax in one calculation. That combination shapes which materials a copy drafting task must retrieve. Four document types carry the relevant wording: the exemption, offset and refund declaration detail table; export contracts; the sales detail ledger; and export customs declarations. These materials update monthly and originate from the electronic tax bureau, ERP sales records, customs electronic data and financial ledgers.
Two features of these materials directly affect retrieval. First, the product name, quantity and trade terms on an export contract and on a customs declaration are different fields of the same transaction. A sentence drafted from one document cannot be extended to another document without checking whether that other document carries the same field. Second, the period scope of the declaration detail table and the sales detail ledger must match. Any sentence about a period, quantity or amount needs to be checked against the same period and the same field. Retrieval returns reference content for a drafter to check. Retrieval does not decide authorization or entitlement.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Candidate bases that use different vector models cannot be associated together. When a list of candidate bases is available, compare the vector model identifier of each base, exclude those that do not match, and record what was observed. Do not run a cross-model search to compare missed recalls.
Search modes include semantic search, full-text search and hybrid search. Compare these modes against each test question. No single mode is assumed to suit all materials. Hybrid search merges multiple recall paths through RRF. RRF is a fusion process, not a separate switch that can be turned on. Result reranking requires a rerank model and must be enabled in the settings. Image retrieval requires image input and matching model capability. Optional paths shown in the overall flow diagram do not all execute on every request.
The reference limit is measured in tokens. The minimum relevance score takes effect only in semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion. Semantic recall and relevance scores do not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to business registration sites or other external systems. The node outputs an array of references whose length can be zero, and the downstream path still runs when no content is found.
Hypothetical paper test case
Material text one (export contract excerpt): Party A is a domestic producer enterprise, Party B is an overseas buyer, the contract subject is self-produced A-type motors, quantity one thousand units, trade terms FOB, contract signing date in March of a given year, and the agreed shipment period is April of the same year.
Material text two (export customs declaration excerpt): The domestic consignor is the producer enterprise named above, the product name is A-type motors, quantity one thousand units, trade terms FOB, and the declaration date is April of the same year.
Material text three (exemption, offset and refund declaration detail table excerpt): The period covered is April of the same year, the export sales amount column shows the amount corresponding to A-type motors, and the remarks column is blank.
Draft sentence to be assessed: This enterprise has obtained the export tax refund for its A-type motors, supply is stable, and long-term large-volume supply is possible.
Handover request: Retrieve the three materials above and determine whether the phrase "has obtained the export tax refund" in the draft sentence is supported by the materials.
Configuration and implementation checks
| Configuration or check item | Approach for this page | Basis or boundary |
|---|---|---|
| Associated knowledge bases | Associate the bases holding the export contract, export customs declaration, declaration detail table and sales detail ledger | Must use the same vector model; check identifiers and exclude non-matching bases |
| Search mode | Test semantic, full-text and hybrid modes separately with the draft sentence and material sentences | No single mode is assumed to suit all materials |
| Reference limit | For this hypothetical trial, start at a value covering one contract plus one customs declaration, and compare with a value covering only a single document | Unit is tokens; the starting value is hypothetical for this trial, not an official default |
| Minimum relevance score | Label samples as relevant or irrelevant to the draft sentence first, then adjust the threshold | Takes effect only in semantic search or when result reranking is used |
| Question optimization | Compare the original question with the optimized question and check whether the subject was altered | Used for coreference resolution and question expansion |
| Result reranking | Compare ordering after a rerank model is available and enabled in settings | Optional path; does not execute on every request |
Three common errors
- Treating a retrieved material as a basis for permission or identity. Retrieval returns reference content only. It does not determine whether a subject may view a record or holds a qualification. If the draft sentence concerns a subject qualification, a separate business verification is required, and the retrieval result alone does not support it.
- Reading "has been shipped" or "has been declared" as "has obtained the export tax refund." In material text three above, the remarks column is blank and does not state a refund outcome. That field remains unsupported. Record the evidence gap rather than writing it as a denial or a conflict.
- Mixing wording from different periods or different fields. The contract agrees a shipment period of April, the customs declaration date is April, and the declaration detail table covers April, so the periods match. If one material covered a different period, it could not be joined with the others into a single conclusion.
How to verify the setup
- Input the three material texts and the draft sentence above, then compare "has obtained the export tax refund" in the draft sentence with the remarks column of material text three. If the remarks column remains blank and no other material states a refund outcome, the phrase is unsupported and the gap must be recorded. Only if a material clearly states the refund is complete, with matching subject, period and field, may the phrase be judged supported.
- Input a clearly labelled paper empty-result array [] and check whether the downstream design runs when the reference content length is zero. Do not claim that a fictional rule search must return [].
- Input the candidate base list and compare the vector model identifier of each base. Exclude and record bases whose identifiers do not match. Only bases with matching identifiers may be associated. Do not run a cross-model search to compare missed recalls, and do not infer run results after joint association.