What the documents look like
Producer export tax refunds combine the offset of domestic sales tax with the refund calculation, so a knowledge base for this work has to hold both filing-perspective and performance-perspective materials. The everyday documents are the exemption, offset and refund filing detail schedule, export contracts, the sales detail ledger, and export customs declarations. Data arrives from the electronic tax bureau, ERP sales records, customs electronic information, and the financial ledger, and it updates monthly. In order fulfilment and after-sales question handling, sales documentation staff and tax accountants raise reconciliation requests about the declared customs value, the contract terms, and the booked amount for a given export shipment. Retrieval has to organise these materials by the same entity, the same period, and the same field so that a later human judgement can be made. The retrieval result supplies quoted content only; it does not carry authorisation, approval, or database writes.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be linked together. The check is to confirm the vector model identifier first, exclude those that do not match, and record what was actually observed; a cross-model search is not run in order to compare missed recalls. Retrieval treats the knowledge base as the smallest search unit, and a collection serves only as a classification grouping.
Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, and hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process in the hybrid path and is not a switch. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold; it does not constitute a permission, entity, or business isolation boundary. The quotation limit is measured in tokens. The output quotation is an array whose length can be zero, so the downstream chain still runs when nothing is found. Question optimisation performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires image input and the corresponding model capability. Reranking requires a model and is enabled according to configuration. The paths shown in the overall flow diagram do not mean that every request executes all of them.
Configuration and implementation checks
Hypothetical paper exercise sample: a tax accountant at an export production enterprise submits a reconciliation request, and the attached source text consists of three sentences. Sentence one: "In March 2026 this enterprise exported a batch of self-produced goods, export contract number H-2026-03, with a contract FOB price of USD 120,000." Sentence two: "The export customs declaration for the same month states an FOB price of USD 118,500 for this batch." Sentence three: "The sales detail ledger for the same month records a booked amount of USD 118,500 for this batch." The draft sentence to be assessed is: "The contract price, the customs declared price, and the booked amount for this export are consistent and can be entered directly into the exemption, offset and refund filing detail schedule." The judgement requires the same entity, the same period, and the same field. The contract price and the customs declared price are different fields, and the third sentence supports only that the customs declared price and the booked amount share a value.
| Configuration item | Relevant setting or check on this page | Correspondence with the sample |
|---|---|---|
| Linked knowledge bases | Link one or more knowledge bases that use the same vector model | Store the filing detail schedule, export contracts, sales detail ledger, and export customs declarations separately, and check that the vector model identifiers match |
| Search mode | Compare semantic, full-text, and hybrid retrieval against the test question | Contract number H-2026-03 is an exact identifier, so check it with full-text or hybrid retrieval |
| Quotation limit | Unit is tokens; test content completeness with actual materials of differing length | For this exercise the starting value is set at 2000 tokens, because the batch materials contain two longer passages, the contract and the customs declaration; compare against a value of 1000 tokens to see whether the quotation is truncated |
| Minimum relevance | Takes effect only under semantic retrieval or when result reranking is used | Label relevant and irrelevant samples first, then adjust the threshold; it is not used to exclude unrelated business |
| Question optimisation | Used for coreference resolution and question expansion | Compare the original question with the optimised question and check whether the entity was altered |
| Result reranking | Requires a model and is enabled according to configuration | When candidates are numerous, check the reranking score; no ordering outcome is presumed |
Three common errors
- Treating the contract price and the customs declared price as the same field and declaring them consistent. In the sample the contract states USD 120,000 and the customs declaration states USD 118,500. These are different fields, so no contradiction conclusion follows; the difference is recorded for later verification.
- Using the minimum relevance value to filter out unrelated business or unauthorised records. That value is a filter threshold, and retrieval does not carry authorisation. Whether an unauthorised record sits inside the knowledge base belongs to another step, and this page makes no assertion about it.
- Treating a retrieved quotation as a completed approval or write. The quotation output chain still runs when its length is zero, and a retrieval result does not constitute authorisation, approval, or a database write.
How to verify the setup
- Input sentence one and sentence two of the sample, compare the values of the two fields, the contract price and the customs declared price, and observe whether the retrieved quotation presents both source passages. If only one passage appears, record the gap and do not judge consistency or contradiction.
- Input the draft sentence, compare its "consistent" wording against the three source sentences, and observe whether a mutually exclusive value appears for the same entity, the same period, and the same field. The sample does not state that the contract price and the customs declared price are comparable, so the finding is unsupported.
- Input an explicit paper input of an empty array, check how the downstream design handles an empty quotation, and observe whether the chain runs. Do not claim that a fabricated rule search necessarily returns an empty array, and do not guarantee a first hit or a complete recall.