What the documents look like
Export tax rebate filing work happens after goods leave the country, and it matches tax vouchers to that export. The main data sources are the electronic tax bureau, the export rebate filing system, customs electronic messages and the finance ledger. Store service and internal training for this category revolve around four material types: the export tax exemption and rebate filing form, the export customs declaration, the purchase invoice and the foreign exchange receipt voucher.
These four materials carry different subjects, objects and status fields. A customs declaration states the exporter and the commodity information. A purchase invoice states the supplier and the goods. A foreign exchange receipt voucher states the fact of receipt. A filing form states the filing period and the rebate amount. A retrieval that hits only one of these types does not show that the voucher chain is complete. Data updates monthly and per batch; this page does not depend on that rhythm, so it is not expanded here.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Before retrieval, check the vector model identifier of each base, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
Semantic retrieval relies on vector distance. Full-text retrieval relies on keyword hits. Hybrid retrieval merges the two paths through RRF. Rerank performs a second ordering of candidate text. RRF is a fusion process, not a switch that can be turned on. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold, not a sorting rule, and it does not carry permission or entity isolation. The reference cap is measured in tokens, and the output reference array can have a length of 0. Question optimization handles coreference resolution and question expansion and adds one model call. Image retrieval requires the corresponding image input and model capability, and reranking requires a model and must be enabled in settings. Optional paths shown in the documented flow diagram do not mean every request executes all of them.
Hypothetical paper exercise sample: a store hands over a paper packet whose source text reads: "March 2025 export customs declaration, operating unit Jia Foreign Trade Company, commodity cotton women's shirts, quantity 1200 pieces; corresponding purchase invoice issued by Yi Garment Factory, goods name cotton women's shirts, quantity 1200 pieces; foreign exchange receipt voucher payee Jia Foreign Trade Company, amount consistent with the invoice; export tax exemption and rebate filing form, period March 2025, status column blank." The draft sentence to be judged is: "This export batch has a complete voucher chain and can be submitted for rebate filing." In this sample the filing form status column is blank. Signing, reaching cooperation and contract effectiveness are different facts. A quotation validity period does not indicate a supply deadline. A complete voucher chain and eligibility for submission are both unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only rebate material bases with a matching vector model identifier | Check each base's vector model identifier first, exclude non-matching ones, record the actual observation |
| Search mode | Run semantic retrieval and full-text retrieval once each | Use the same test question to compare the recalled content of the two paths; do not assume one mode suits all materials |
| Reference cap | Starting point set to a token amount covering one customs declaration plus one invoice | Test with materials of different lengths to see whether content is complete, then adjust the starting point |
| Minimum relevance | Starting point set to 0.5 for this exercise | Label relevant and irrelevant samples first, then adjust the threshold and compare results before and after filtering |
| Question optimization | Enabled for short questions in this exercise | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled in this exercise when a rerank model is available | Compare candidate order before and after reranking and confirm the rerank score can be used for filtering |
Three common errors
- Treating a retrieval hit as authorization. The input is the draft sentence from the sample above. If the system recalls a filing form fragment, the wrong action is to conclude from this that submission is allowed. Retrieval does not handle permission or entity isolation, and a blank status column on the filing form is unsupported.
- Associating knowledge bases across vector models. The input is two bases with different vector model identifiers. The wrong action is to associate them directly and then compare recall differences. The correct action is to check the identifiers, exclude the non-matching base and record the actual observation.
- Treating pending review as a lack of operating permission. The input is the blank status column on the filing form in the sample. The wrong action is to decide the store lacks operating permission. Pending review does not show a lack of operating permission; the evidence gap must be recorded.
How to verify the setup
- Input the sample source text and draft sentence above, and compare whether the recalled content includes the filing form status field. If the recalled content covers only the customs declaration, invoice and foreign exchange receipt voucher, the phrase "complete voucher chain" in the draft sentence is unsupported and the gap should be recorded.
- Input the same test question through semantic retrieval and full-text retrieval separately, and compare the returned content of the two paths. If an identifier-type question is hit by full-text retrieval and missed by semantic retrieval, this is an expected phenomenon; do not conclude from this that one mode has failed.
- Input an explicit paper empty result [] and compare whether the downstream design accommodates a reference output of length 0. If the downstream does not define an empty-result branch, record that gap. Do not claim that a fictional rule retrieval necessarily returns [].
This page addresses only export tax rebate filing, store service and internal training, and knowledge base retrieval and recall. At operational handover, which materials, operations and requests sit in which authorization state cannot be treated as authorized execution through retrieval or a form.