What the documents look like
Export tax rebate filing work happens after goods leave the country, when tax vouchers are matched against the export record. The rhythm is monthly or per shipment. The working documents are the export tax exemption and rebate filing form, the export customs declaration, the purchase invoice, and the foreign exchange receipt voucher. A product question in this setting has to connect these documents to each other: the product name and specification model on the customs declaration need to be traceable to the goods name and quantity on the purchase invoice for the same shipment.
Errors in rebate filing lead to delayed rebates, ineligible rebates, or tax risk. The retrieval step has a narrower job. It returns knowledge base reference content that relates to the current product, batch, or voucher so that a filing clerk or a question-answering flow can review it, or so a downstream process can use it. Retrieval does not perform voucher matching and does not replace tax judgment.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. The operator checks the vector model identifier first, excludes those that do not match, and records what was actually observed. A cross-model retrieval is not run in order to compare missed recalls.
Search mode is compared per test question across semantic retrieval, full-text retrieval, and hybrid retrieval. No single mode is assumed to suit all materials. Semantic retrieval depends on vector distance and suits natural-language questions. Full-text retrieval depends on keyword hits and suits numbers, models, and proper nouns. Hybrid retrieval merges the two result paths through RRF; RRF is a fusion process, not a switch that is turned on. Result reranking requires a ReRank model and must be enabled according to the configuration; after reranking, a score between 0 and 1 is produced. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. That value is a filtering threshold, not the final ranking rule, and it does not carry permission or entity isolation. The reference cap is measured in tokens and controls how much content a single search may reference. Question optimization performs coreference resolution and question expansion, and enabling it adds one model call.
Image retrieval requires the corresponding image input and model capability. The optional paths shown in the overall flow diagram do not mean every path executes on every request. Reference content is output as an array, and its length can be 0; when nothing is retrieved, the output path still completes.
Hypothetical paper exercise sample
The following is paper review material and does not involve a real customer or a legitimate identity number.
Material text one (export customs declaration excerpt): "Declaration date March 12, 2024. Goods name: stainless steel vacuum flask. Specification model: 500 ml direct-drink style. Quantity: 2000 pieces. Domestic shipper: a certain foreign trade company."
Material text two (purchase invoice excerpt): "Invoice date March 5, 2024. Goods name: stainless steel vacuum flask. Specification: 500 ml direct-drink style. Quantity: 2000 pieces. Seller: a certain manufacturing plant."
Draft sentence to be assessed: "The customs declaration and the purchase invoice for this batch of stainless steel vacuum flasks agree on goods name, specification, and quantity, so it may enter the rebate filing data generation step."
This draft sentence concerns rebate filing data generation, which is a downstream business action. The retrieval step can only supply reference content; it does not execute generation or filing.
Configuration and implementation checks
| Configuration item | Assumed trial starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One knowledge base whose vector model identifier is V1 | Associate after checking that the vector model identifier matches; when comparing with a base carrying a different model identifier, record only the identifier difference and do not associate them together |
| Search mode | Hybrid retrieval | The sample contains keywords such as "stainless steel vacuum flask" and "500 ml" together with natural-language description; compare hit content against semantic retrieval alone and full-text retrieval alone |
| Reference cap | Assumed starting point of 2000 tokens | The sample contains two source passages, the customs declaration and the invoice; compare against 1000 tokens to see whether the reference content keeps goods name, specification, and quantity complete |
| Minimum relevance | Assumed starting point of 0.3 | First label relevant and irrelevant passages in the sample, then compare whether irrelevant passages are filtered at different thresholds |
| Question optimization | Not enabled in this exercise | The sample question has a clear subject; compare with the enabled case to see whether optimization changes the "stainless steel vacuum flask" subject |
| Result reranking | Not enabled in this exercise | No ReRank model configuration is provided; the enabling condition is that a model exists and is enabled according to the configuration |
Three common errors
- Treating a retrieval hit as an authorization approval. Input the draft sentence and the reference content returned by retrieval, then compare whether the reference content contains an approval, signature, or effective status. Retrieval returns reference content only and produces no authorization conclusion. If the reference content does not state an approval status, record it as unsupported.
- Treating minimum relevance as an entity or permission filter. Input the sample question and an unrelated goods passage, then compare whether the unrelated passage is filtered after the threshold changes. That value filters only low-relevance results; it does not exclude unrelated businesses or records without permission.
- Describing RRF as a switch that is turned on. Input the hybrid retrieval configuration, then compare whether the semantic and full-text result paths are fused through RRF. RRF is a fusion process, has no independent switch, and does not guarantee a first hit or a complete recall.
How to verify the setup
- Input material text one and material text two, then compare whether the reference content returned by retrieval contains "stainless steel vacuum flask," "500 ml direct-drink style," and "2000 pieces" at the same time. If any field is missing from the reference content, record the gap and do not judge the documents as consistent.
- Input the draft sentence, then compare whether the reference content states an authorization or execution status for the rebate filing data generation step. If it does not state one, record it as pending verification and do not claim the generation step may be entered.
- Input an explicit empty-reference review case, using a paper input of an empty array to review how the downstream design handles reference content of length 0. Do not claim that a fictional rule retrieval necessarily returns an empty array.