What the documents look like
Export tax rebate filing happens after goods leave the country, when tax vouchers are matched against export records. The working materials come from the electronic tax bureau, the export rebate filing system, customs electronic messages, and the financial ledger. Day-to-day handling covers the export tax exemption and rebate filing form, the export customs declaration, purchase invoices, and foreign exchange receipt vouchers. Retrieval demand sits in order fulfilment and after-sales question answering: a user asks whether a purchase invoice linked to a given customs declaration has been matched, whether a foreign exchange receipt voucher has been registered, or what status a filing form for a period currently holds. These materials update monthly and per batch. For the same batch of goods, the customs declaration, invoice, and receipt voucher need to agree on the responsible entity, the period, and the field values before a retrieval result can serve as a basis for answering. The supplied material does not describe automatic synchronization between systems. Retrieval only recalls content already stored in the knowledge base; it does not perform authorization or business execution.
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 built on different vector models cannot be associated together; the vector model identifier is checked and non-matching bases are excluded. Search mode is compared across semantic retrieval, full-text retrieval, and hybrid retrieval against the test question, with no assumption that one mode fits all materials. The citation limit is measured in tokens. The minimum relevance threshold takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion. Hybrid retrieval fuses multiple result paths through RRF, and RRF is a fusion process rather than a switch. Image retrieval requires image input and the corresponding model capability. Reranking requires a rerank model and must be enabled in the settings. Optional paths shown in the overall flow diagram do not mean every request executes all of them. Retrieval and similarity scores do not constitute permission or entity isolation. When no content is retrieved, the output chain still runs through.
Hypothetical drill sample: a finance staff member at a foreign trade enterprise checks the filing materials for one batch of exported goods. A purchase invoice in the knowledge base reads: "Invoice No. FP-2024-0317, issue date 17 March 2024, buyer Company A, goods cotton men's shirts, quantity 1200 pieces, amount 96000 yuan." An export customs declaration reads: "Declaration No. 20240317001, export date 17 March 2024, domestic consignor Company A, commodity cotton men's shirts, quantity 1200 pieces." A foreign exchange receipt voucher reads: "Receipt date 2 April 2024, payer Company B, amount 96000 yuan, corresponding invoice No. FP-2024-0317." The draft sentence to be judged is: "This batch of goods has completed rebate filing and received the rebate payment." The objects of comparison are the three materials above and the draft sentence. The period is 17 March 2024 to 2 April 2024. The authorization boundary is that the finance staff member consults materials already stored for this enterprise. Retrieval produces no filing action and no payment action.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases using the same vector model; check the vector model identifier | The sample materials must sit in one same-model knowledge base; cross-model bases are excluded |
| Search mode | Compare semantic, full-text, and hybrid retrieval against the test question | Invoice and declaration numbers are exact identifiers, suited to full-text or hybrid retrieval |
| Citation limit | Measured in tokens; test content completeness with actual materials of differing length | The three sample materials differ in length; observe whether the citation covers entity and quantity |
| Minimum relevance | Hypothetical trial starting value 0.3, chosen because the sample question carries an explicit number; compare against 0.5 to observe recall changes | Takes effect only under semantic retrieval or result reranking |
| Question optimization | Compare the original question with the optimized question; check whether the entity was altered | The sample question contains an invoice number; observe whether optimization preserves the number |
| Result reranking | With a rerank model present and enabled in settings, filter by the rerank score | Observe ordering when the sample has many candidates; if not enabled, record that it was not enabled |
Three common errors
- Treating a retrieval result as an authorization basis. The input is the sample draft sentence; the comparison objects are the three material texts. The materials carry no field stating that filing was completed or that the rebate payment arrived. A retrieval hit does not mean the business action was executed; record the evidence gap.
- Associating knowledge bases built on different vector models. The input is two knowledge bases whose vector model identifiers differ; compare whether the identifiers match. When they do not match, exclude the base. Do not run a cross-model search to compare missed recall.
- Treating the minimum relevance threshold as a permission or entity filter. The input is the sample question; the comparison object is the recalled content. A similarity score does not exclude unrelated enterprises or unauthorized records. Retrieval does not perform entity isolation.
How to verify the setup
- Input the three sample material texts and the draft sentence. Compare whether the draft sentence shares the same entity, the same period, and the same field as the materials. The materials state no completed filing and no received rebate payment, so the sentence is judged unsupported and pending verification.
- Input invoice number FP-2024-0317 as the retrieval term. Compare whether the recalled content contains that invoice text; observe whether it is matched and check the entity and quantity fields. A first-position hit or complete recall is not guaranteed.
- Input an explicit paper empty result [] and check the downstream design. Observe whether the downstream continues to run through with an empty citation. Do not claim that a fabricated rule makes retrieval always return [].