What the documents look like
Producer export tax exemption, offset and refund combines the offset of domestic sales tax with the refund calculation. Product reference questions therefore draw on four document families at once: the exemption, offset and refund declaration detail sheet, the export contract, the sales subsidiary ledger, and the export customs declaration. These materials update monthly, and their field conventions are spread across the electronic tax bureau, ERP sales data, customs electronic messages and the financial books. Product code, export date, free-on-board price, levy rate and refund rate recur across all four families, so one product can appear as different fields in different files. The vector model and index step takes these materials after chunking and storage, and its job is to let a product reference question recall the corresponding source text for later human checking.
Two features of this category shape that step. First, the same product is described by repeated but not identical fields, so an index built from one field alone may not carry the wording a question actually uses. Second, the declaration detail sheet and the customs declaration both carry the export date and the free-on-board price, and a question can be answered only when the recalled passage and the questioned claim refer to the same subject, the same period and the same field.
Constraints at the vector indexes step
Knowledge base search operates on the whole library as the smallest search unit; a collection only classifies data and does not change search results. One source data block can be linked to several indexes, and a hit on any one index recalls that source block. Knowledge bases attached to the same search node must use the same vector model; knowledge bases with different vector models cannot be attached together. Index content is more useful when complete and focused, and retrieval coverage against index length has to be weighed against the sample at hand. Image retrieval also depends on whether an image description or image vector index was generated at ingestion time, not only on what is submitted at query time.
Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, hybrid retrieval merges multiple result paths through RRF, and ReRank reorders candidates and returns a 0-1 score. The minimum relevance value applies only under semantic retrieval or when result reordering is enabled, and the reference limit is counted in tokens. Retrieval itself does not perform permission or entity isolation.
Hypothetical drill sample. The June 2024 exemption, offset and refund declaration detail sheet of a producer enterprise contains one line reading: "Product code 8501.10.00, export date 2024-06-18, free-on-board price USD 120,000, levy rate 13%, refund rate 13%, self-produced export." The export contract of the same month states: "The buyer pays a 20% deposit 30 days before shipment, and the balance within 15 days after sight of the bill of lading copy." The draft sentence to be judged is: "This product exported in June applies a 13% refund rate, the deposit has been received, and the exemption, offset and refund may be declared." The sentence carries subject, product code, period and refund rate, but no receipt record for the deposit appears in either the detail sheet or the contract, so that element is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Vector model | One model, text input | All four document families are text; compare model identifiers before attaching |
| Search mode | Hybrid retrieval | Product code is an exact query and product description is a semantic query; both paths run together |
| Reference limit | 3000 tokens | A detail sheet line is short; compare recall completeness against a 1500-token starting point |
| Minimum relevance | 0.35 | Compare against a 0.5 starting point and observe whether unrelated fragments decrease |
| Question optimization | Enabled | Product questions often contain references such as "this model"; compare against the disabled state on follow-up turns |
| Index content | Product code, export date, free-on-board price, refund rate | Build one index per detail sheet field and compare single-field indexes against a merged index |
The values above are hypothetical trial starting points for this page, not official defaults. RRF is the result fusion process of hybrid retrieval and is not configured as a separate switch.
Three common errors
- Attaching knowledge bases that use different vector models. Input a retrieval request for "product code 8501.10.00", compare the model identifier of each library, and exclude any library whose identifier differs. Do not run a cross-model retrieval to infer missed recall.
- Treating a semantic similarity score as an authorization decision. Input the draft sentence "the exemption, offset and refund may be declared", compare the retrieval result against the authorization record, and note that a retrieval hit does not mean declaration permission has been approved. Authorization status has to be checked separately.
- Reading "deposit received" out of the contract payment clause. Input the contract wording "a 20% deposit 30 days before shipment", compare the sales subsidiary ledger and the bank receipt record, and treat the contract term and the arrival of funds as two separate facts. With no receipt record, the status remains unsupported.
How to verify the setup
- Input the detail sheet line and the draft sentence from the hypothetical sample, and compare each field of the draft sentence against the corresponding field of the source. The source carries product code, export date, free-on-board price, levy rate and refund rate; the draft sentence's "deposit has been received" has no corresponding record and is judged unsupported, while the 13% refund rate matches the source and is judged supported.
- Input "product code 8501.10.00" as an exact query and compare the source lines returned by full-text retrieval and by semantic retrieval. If full-text retrieval returns the line and semantic retrieval does not, record that observation. If neither path returns it, record the empty result, and use an explicit paper input of an empty array to check whether the downstream reference path still completes.
- Input "what is the refund rate for this model" as a follow-up turn and compare the recalled source lines with question optimization enabled and disabled. If the reference is completed to a specific product code and the corresponding line is recalled, record that observation. If unrelated lines are still recalled, record that the expectation was not met and do not conclude from it that the model has failed.