What the documents look like
Producer export tax exemption, offset and refund combines domestic sales tax offset with refund calculation. The declaration detail schedule, export contract, sales subsidiary ledger and export customs declaration must agree on the same entity, the same period, the same product and the same amount basis. For product data and selection questions, an answer has to point back to a specific location in these materials.
The typical working parties are export producers, tax accountants and sales document staff. Materials update monthly. A citation trace must point to the currently effective version; an earlier version serves only as a comparison point.
Three industry facts shape this task. First, domestic offset and refund are calculated together, so a cited amount cannot be read in isolation from the offset computation. Second, the four document types must stay consistent on entity, period, product and amount, so a citation that matches only one field is incomplete. Third, the declaration detail schedule is the document that carries the final claim, so a draft sentence that omits its corresponding fields remains unsupported.
Constraints at the citation tracing step
The knowledge base search node can link one or more knowledge bases that use the same vector model. It outputs cited content as an array, and the length may be zero. An empty array still passes through the output path.
After retrieval results reach the answering step, each answer sentence must be checked line by line against the cited source text. The minimum relevance value takes effect only under semantic retrieval or result reranking. It is a filter threshold and does not perform permission or entity isolation. Retrieval itself is not authorization.
Hypothetical exercise sample. A producer's tax accountant is organizing materials during the July 2025 filing period. The source text reads: "Export contract HT-2025-0712 specifies 100 units of Model A motor, FOB total RMB 580,000, contracting entity Company A." "Sales subsidiary ledger for June 2025 shows Model A motor export revenue RMB 580,000, booking entity Company A." "Export customs declaration number 20250703 shows 100 units of Model A motor, domestic consignor Company A." The draft sentence to be judged is: "Company A's 100 units of Model A motor export revenue has been checked and agrees, and can be used directly for the exemption, offset and refund declaration." This sentence lacks the correspondence between the customs declaration amount field and the declaration detail schedule, so it is unsupported.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select only exemption, offset and refund material bases with the same vector model | Compare identifiers against a base using a different vector model; if identifiers differ, do not link them together |
| Search mode | Hybrid retrieval | Compare against semantic retrieval results for how numbering and model-type hits appear |
| Minimum relevance | Hypothetical starting value 0.3 | Compare retrieved text line by line against source text; observe whether unrelated fragments remain |
| Citation upper limit | Hypothetical starting value 2000 tokens | Compare against declaration detail schedule field completeness; observe whether key fields are cut off |
| Citation output handoff | Citation array passed to the answering node | Check entity, period and field between the answer sentence and the cited source text |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare retrieval terms against the disabled state; observe whether references are completed |
These values are starting points for this exercise, not official defaults. The comparison method defines how each starting point is tested against observed material.
Three common errors
- Treating a retrieval hit as authorization. A non-empty citation array only means content was found; it does not mean the material has been reviewed or can be used for a declaration. Check input is the three source sentences above. Compare cited content against declaration detail schedule fields. Observe whether an amount field has no correspondence.
- Treating minimum relevance as a permission filter. This value takes effect only under semantic retrieval or result reranking. It does not exclude unrelated businesses or unauthorized records. Check input is a cross-entity query. Compare returned citations against the retrieval scope. Observe whether records from another entity appear.
- Treating an empty citation array as a workflow error. The documentation states the length may be zero and the output path still runs. Check input is a paper empty array. Compare the downstream node design. Observe whether the downstream handles the empty result branch.
How to verify the setup
- Input the three source sentences and the draft sentence given above. Compare entity, period and field between the draft sentence and the source text. The citation support condition is met only when all three fields correspond; otherwise record the result as unsupported.
- Input the identifiers of two knowledge bases using different vector models. Compare whether the identifiers match. If they differ, exclude that base and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
- Input a paper empty array. Compare whether the downstream answering node handles the empty result. Observe whether the output path runs. This check does not assert that a fictional rule retrieval always returns an empty array.
The checks above state expected conditions. They do not claim that any test, deployment, approval or declaration action has taken place. Where a field is not provided in the source text, the value remains unknown.