What the documents look like
Export tax rebate filing happens after goods leave the country, and it matches tax vouchers to that export. In the inventory procurement and supplier records step, staff align purchase invoices, export customs declarations, foreign exchange receipts and the export tax exemption or rebate filing form item by item. These materials update monthly or per batch, and their fields include supplier name, commodity code, quantity, amount, invoice number and customs declaration number. The subject, period and fields across vouchers must agree. When retrieval mixes fragments from different suppliers or different batches, later matching loses its verifiable basis. The vector model and index task makes these materials semantically retrievable inside a knowledge base. Retrieval itself does not perform authorization, review or filing submission.
Constraints at the vector indexes step
Knowledge bases linked for retrieval must use the same vector model. Check the actual model identifier and input type first; knowledge bases with different vector models cannot be linked together. One source data block can carry several indexes, and the index content and the original text are checked separately. Index content benefits from being complete and focused, while retrieval coverage and index length are weighed against samples. Image retrieval depends on whether image descriptions or image vectors were generated at ingestion time; uploading an image only on the query side does not create a hit condition. After a model or index change, compare cited original text using the same question set, and do not promise an effect from a vector score alone. The minimum relevance value is a filter threshold, and semantic recall and relevance do not enforce permission or entity isolation; retrieval does not constitute authorization.
Hypothetical drill sample: a finance staff member at a foreign trade enterprise checks a batch of purchase invoices and customs declarations in the inventory procurement and supplier records knowledge base. The three source sentences are: "Supplier A issued a purchase invoice in March 2025, invoice number A001, commodity code X, quantity 100 units, amount 10,000 yuan." "The corresponding export customs declaration number is B001, declaration date March 2025, export quantity 100 units." "Foreign exchange receipt C001 shows 10,000 yuan received in April 2025." The draft sentence to be judged is: "This batch of goods from Supplier A has completed the tax rebate filing and been approved." The source text records only invoice, customs declaration and foreign exchange receipt information. It does not record filing submission, review or approval status, so the draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Linked knowledge base | Inventory procurement and supplier records knowledge base | Check each base's vector model identifier first, exclude inconsistent models, then link |
| Search mode | Hybrid search | Exact fields such as invoice numbers and customs declaration numbers use full-text recall, natural language descriptions use semantic recall, merged through RRF |
| Minimum relevance | Hypothetical starting point 0.3 for this trial | This value is a hypothetical starting point; adjust after comparing cited original text for the same question set at different values |
| Citation limit | Hypothetical starting point 2,000 tokens for this trial | This value is a hypothetical starting point; adjust after comparing cited original text length and question coverage |
| Result reranking | Enable ReRank | When candidates are many, rerank text and filter by rerank score; confirm a rerank model is configured |
| Question optimization | Enable only for multi-turn follow-ups | Question optimization calls the model again; use it for short or ambiguous questions, and decide after comparing response time |
Three common errors
- Linking knowledge bases that use different vector models. The input is two bases with different model identifiers. Compare the model identifier fields. When they differ, exclude the base and do not run cross-model retrieval to compare missed recalls.
- Treating a retrieval hit as an approved rebate filing. The input is the draft sentence "has completed the tax rebate filing and been approved." Compare it with the source text. The source does not record filing or approval status, so the judgment is not supported, and the evidence gap remains.
- Treating minimum relevance as a permission or entity filter. The input is records from different suppliers. Compare field values. Relevance affects similarity filtering only and does not exclude unrelated business or unauthorized records; authorization requires a downstream system implementation.
How to verify the setup
- Enter the three hypothetical source sentences and the draft sentence. Compare whether the draft sentence agrees with the source text on subject, period and fields. When the source does not record filing and approval status, judge it as not supported, not as a denial or conflict.
- Enter invoice number A001. Compare whether the full-text recall result contains that original text block, and observe whether the cited original text corresponds to the entered number. When there is no hit, record the gap; do not guarantee a first-position hit or full recall.
- Enter a paper question with no corresponding material. Compare the output citation array. If it is [], check how the downstream design handles empty results; do not claim that a fictional rule search always returns [].