What the documents look like
Export tax rebate filing happens after goods leave the country, and the work is matching tax vouchers. The core materials are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices, and foreign exchange receipt vouchers. Contract and transaction material comparison requires that the invoice, customs declaration, and receipt record for the same entity corroborate one another on commodity name, quantity, amount, and period. Filing materials are gathered monthly or per batch, and document clerks and tax accountants complete the material check before filing. These traits shape retrieval input: it often carries precise identifiers such as declaration numbers, invoice numbers, and commodity codes, mixed with natural language descriptions, so a retrieval configuration has to serve both identifier queries and semantic queries.
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; the vector model identifier is checked first and non-matching bases are excluded, with the actual observation recorded. No inference is drawn about what cross-model association would produce at runtime.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval fuses multiple result paths through RRF; RRF is a fusion process, not a switch that can be turned on by itself. The citation limit controls, in tokens, how much content each search cites. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval.
Hypothetical paper exercise sample
A document clerk at a foreign trade enterprise gathers January 2026 export batch materials in March 2026. The source text for checking reads: "Export customs declaration number A20260115, export date 15 January 2026, commodity name cotton men's shirts, quantity 1200 pieces, total price USD 24,000. Corresponding purchase invoice number FP20260108, invoice date 8 January 2026, commodity name cotton men's shirts, quantity 1200 pieces, amount RMB 168,000. Foreign exchange receipt voucher SW20260220, receipt date 20 February 2026, amount USD 24,000."
The draft sentence to be assessed is: "The customs declaration, purchase invoice, and receipt voucher for this batch agree in quantity and correspond in amount, so document matching can be confirmed as complete." This draft sentence concerns a document-matching fact outside signing, cooperation reached, or contract effectiveness. The authorization boundary is material checking only, and does not include filing submission or rebate confirmation.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one rebate material base with the same vector model identifier | Check each base's vector model identifier first, exclude non-matching ones, and do not run a cross-model retrieval comparison |
| Search mode | Hybrid retrieval | Declaration and invoice numbers need full-text hits, commodity name descriptions need semantic recall; compare the candidate results each mode returns |
| Citation limit | Hypothetical starting value of 3000 tokens | This is a paper exercise starting value, used to compare whether material fragments of different lengths enter the cited content in full, then adjusted against actual observation |
| Minimum relevance | Hypothetical starting value of 0.4 | This value takes effect only in semantic retrieval or result reranking; label relevant and irrelevant samples first, then compare how the filtered candidates change |
| Question optimization | Enabled for follow-up questions containing pronouns | Compare the original question with the optimized question, check whether the entity or object was altered, and keep it off for already explicit questions |
| Result reranking | Reranking model enabled according to settings | Reranking requires the model and must be enabled according to settings; compare candidate ordering and scores before and after reranking |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input the declaration number of a different entity from the hypothetical sample, compare whether the returned content still contains that entity's materials, and observe whether the result is decided by similarity alone. A relevance threshold does not isolate unrelated enterprises or unauthorized records, and retrieval by itself is not authorization.
- Associating knowledge bases with different vector models directly. Input two bases with different vector model identifiers, compare whether the association operation is permitted, and observe whether cross-model retrieval behavior appears. The correct step is to check identifiers, exclude non-matching ones, and record the actual observation without inferring runtime results.
- Treating a pending or draft sentence as a completed fact. Input the draft sentence above, compare its wording with the source text for checking, and observe whether the source text states that document matching has been confirmed. The draft sentence is a pending statement and does not replace a checking conclusion.
How to verify the setup
- For declaration number A20260115 in the hypothetical sample, input that number and run retrieval, compare whether the returned content contains the corresponding declaration source text, and observe whether the commodity name, quantity, and total price in the cited content match the source text. If the return is empty, record the empty result and check whether the data has finished training; do not claim that an identifier search necessarily hits.
- For the draft sentence "The customs declaration, purchase invoice, and receipt voucher for this batch agree in quantity and correspond in amount, so document matching can be confirmed as complete," input that sentence and run retrieval, compare the returned content with the source text for checking, and observe whether the source text states a matching confirmation status. The draft sentence is supported only when the source text states a confirmed fact for the same entity, same period, and same field; when the source text does not state it, record the item as pending verification.
- For receipt voucher SW20260220 in the hypothetical sample, input that voucher number with reranking enabled, compare candidate ordering and scores before and after reranking, and observe whether the rerank score is used for filtering. If the reranking model is not enabled, record that this path was not executed and do not count an unexecuted path in the verification conclusion.