What the documents look like
Export tax rebate filing takes place after goods have left the country, and it matches tax vouchers against what was declared. The main data sources are the electronic tax bureau, the export rebate filing system, customs electronic messages and the accounting ledger. The working documents are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices and foreign exchange receipt vouchers. Finance staff at trading companies, tax accountants and document clerks handle this work. Errors lead to delayed rebates, rebates that cannot be granted, or tax risk. These materials update monthly or per shipment. Citation tracing requires that every conclusion in an answer can be traced back to its source text, with the version and scope of that source stated.
For this page the task is order fulfilment and after-sales question answering, and the feature under examination is citation source and tracing. The relevant unit of search is the knowledge base as a whole; a collection only groups data and does not change search results. A search node returns citations as an array, and that array may have a length of zero, meaning the output path still completes even when nothing was found. The citation cap is expressed in tokens, not in a top-k item count. The minimum relevance value is a 0–1 filter threshold that applies only under semantic search or when result reranking is used; it filters low-relevance results and does not act as a permission or entity filter. Hybrid search merges multiple recall paths through RRF, which is a fusion process rather than a separate switch. Knowledge bases may be linked together only when they share the same vector model. Question optimization adds one model call before retrieval for coreference resolution and question expansion. Rerank performs a second ordering of candidate text and returns a 0–1 score. Image retrieval requires image input and matching model capability, and reranking requires a model and must be enabled by configuration. Optional paths drawn in a document flow diagram do not mean that every request executes all of them.
Constraints at the citation tracing step
The search node outputs a citation array whose length may be zero, and the downstream path continues in that case. The citation cap limits how many tokens of content may be cited per search. The minimum relevance value takes effect only under semantic search or result reranking and is a filtering threshold, not a final ordering rule and not an authorization boundary. RRF fusion means a result hit by several recall paths is usually ranked earlier, so final citation order does not follow a single vector similarity. Linked knowledge bases must share the same vector model; libraries with different vector models cannot be linked together. Question optimization adds a model call before retrieval, which improves multi-turn retrieval but adds latency. Rerank reorders candidates and returns a 0–1 score that can be used for filtering. Image retrieval needs image input plus model capability, and reranking needs a model that is enabled by configuration.
Hypothetical paper exercise sample: a finance clerk at a trading company is handling an after-sales question during order fulfilment and receives an internal draft sentence reading "This export transaction already meets the conditions for filing a rebate and may be submitted." The source materials supplied for checking are as follows. First, the export customs declaration states an export date of 10 March 2025, goods of textiles, quantity 500 pieces. Second, the purchase invoice states an invoice date of 5 March 2025, with an amount corresponding to the customs declaration. Third, the foreign exchange receipt voucher states a receipt date of 2 April 2025. These materials do not state any review status or authorization conclusion for the rebate filing. This sample is used for paper checking only and does not represent any real business or actual filing.
Configuration and implementation checks
Configuration here covers only settings documented in the official material and implementation checks for this page. The values below are starting points for this hypothetical exercise, with the reason for choosing them and the value they are compared against.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select one knowledge base with a consistent vector model | Knowledge bases may be linked only when the vector model is the same; check the identifier first and exclude those that do not match |
| Search mode | Hybrid search | Uses semantic and full-text recall together and merges them through RRF, suited to a mix of identifiers and natural-language questions |
| Citation cap | Set to a token amount that can hold the three source texts | Compare against the total length of the source texts so the citation covers the customs declaration, invoice and receipt voucher |
| Minimum relevance | Set to 0.5 | Compare against the 0–1 rerank score to filter low-relevance fragments; not used to exclude unrelated businesses or unauthorized records |
| Result reranking | Enabled | Performs a second ordering when candidates are numerous; compare ordering stability against results without reranking |
| Question optimization | Enabled only for multi-turn follow-ups or short questions | Compare retrieval latency and recalled content against the disabled state; it can be disabled when the current question is already clear |
Three common errors
- Treating the minimum relevance value as a permission or entity filter. Input is the sample retrieval above; compare the retrieval results against the authorized scope and observe whether unrelated companies or unauthorized records appear. A relevance score does not perform permission filtering, so the appearance of unrelated records means a separate permission control is needed and cannot be resolved by raising or lowering relevance.
- Treating retrieval output as filing authorization. Input is the draft sentence and the sample materials; compare whether the materials state a review status or authorization conclusion, and observe whether the draft sentence claims the filing conditions are met. When the materials state no review status, the draft sentence is unsupported, and retrieving materials does not mean the filing has been authorized.
- Linking knowledge bases with different vector models. Input is two knowledge bases with different vector model identifiers; compare the model identifiers and observe whether they are allowed to be linked. When the identifiers differ they should be excluded, and no cross-model retrieval is performed to compare missed recall or redundancy.
How to verify the setup
- Citation completeness check. Input is the sample source texts and the draft sentence; compare whether the citation array contains all three source texts, the customs declaration, the invoice and the receipt voucher, and observe whether the cited content covers the export date, quantity, amount and receipt date. If any material is missing from the citation array, record the gap and do not conclude that the filing conditions are met.
- Empty-result handling check. Input is a paper citation array of []; compare the design of the downstream answering step and observe whether the downstream path continues and states that no content was retrieved. An empty array does not mean the workflow errored out, and it is not claimed that a fabricated rule search must return [].
- Version and status check. Input is the sample materials and the draft sentence; compare the dates, quantity and amount stated in the materials against the filing conditions claimed in the draft sentence, and observe whether there are mutually exclusive values for the same subject, the same period and the same field. When the materials state no review status or authorization conclusion, record this as pending verification and preserve the evidence gap without supplying an authorization or effective clause.