What the documents look like
Producer export tax refunds offset domestic sales tax against export refunds in a single calculation. Inventory procurement and supplier records therefore carry two jobs at once: showing that the raw materials behind exported goods were genuinely purchased, and supporting the non-exempt, non-deductible amount inside that combined calculation. Materials handled month by month include the refund declaration detail schedule, export contracts, the sales detail ledger and export customs declarations; the procurement side adds purchase contracts, input invoices and goods receipt notes. The same shipment can appear in a purchase receipt record, the sales detail ledger and a customs declaration at the same time, so chunking has to keep the fields describing one business fact inside a range that can refer back to each other.
Constraints at the document chunking step
Three constraints shape this step. Procurement records and export records must retain a document identity and version that can be traced back, otherwise it is impossible to judge whether a given input item corresponds to a given export item. Segmentation has to follow the structure of the specific material, and conditions, notes and the content they govern cannot be separated; quantity, unit price, tax rate and validity range in a purchase contract belong in the same chunk. Original data and retrieval indexes are distinct: one group of original data can map to several indexes, the vector score takes the highest score, retrieval recalls vectors first and then uses the vector ID to find the original data in MongoDB, and entries pointing to the same original data are merged.
Hypothetical paper exercise sample: an exporting producer holds a purchase contract with three original sentences. "Contract number CG-2025-0417, supplier is a domestic materials plant, subject matter is 120 tonnes of cold-rolled steel coil, tax-inclusive unit price 4,300 yuan per tonne, delivery date May 2025." "All goods under this contract are used in the production of export order EX-2025-0088." "This contract takes effect on the date both parties sign and remains valid until 31 December 2025." A draft sentence awaits judgment: "This supplier record can be used for input aggregation in the current refund declaration." The fields needed for that judgment are supplier name, subject matter, quantity, corresponding export order number and contract validity period. The contract does not state the supplier taxpayer identification number or an invoice number; both remain unsupported and must be recorded as evidence gaps.
Configuration and implementation checks
The table lists only settings or implementation checks already described in the official material and relevant to document parsing and chunking on this page. The values are starting points for this hypothetical trial, used to compare against another value; they are not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Purchase numbers and contract numbers are exact queries, while natural-language questions are semantic queries; both question types need coverage, compared against semantic retrieval alone |
| Citation cap | 2,000 tokens as this trial's starting point | Purchase contract passages are long; compare against a 1,000-token starting point and observe whether quantity and validity fields are truncated |
| Minimum relevance | 0.4 as this trial's starting point | Takes effect only in semantic retrieval or with result reranking; compare against a 0.2 starting point and observe whether unrelated procurement passages enter the citations |
| Result reranking | ReRank enabled | Hybrid retrieval leaves many candidates, and the 0-1 rerank score can be used for filtering; compare against the disabled state for ordering stability |
| Question optimization | Enabled for multi-turn follow-ups | Question optimization adds one model call before retrieval; compare against the disabled state for latency and for how well references are completed |
| Index check | Verify the correspondence between the indexes field and original data | One group of original data can map to several vectors; check whether image or description indexes were generated according to model capability |
Three common errors
- Treating a successful upload as a completed retrieval update. Input is a revised purchase contract, the comparison object is the passage content returned by retrieval, and the observable result is that the returned content is still the earlier version. Record the update status as pending verification.
- Treating a similarity score as a permission or entity filter. Input is a procurement passage query spanning suppliers, the comparison object is the supplier owning each returned passage, and the observable result is that passages from another supplier with the same material name are returned. Record that the retrieval scope was not separated by supplier.
- Treating a pending review request as an approved execution. Input is one input aggregation request, the comparison object is the request status field together with the authorization record, and the observable result is a submission record with no approval record. Judge this as pending verification and do not execute aggregation on that basis.
How to verify the setup
- Input the three original sentences of the purchase contract from the hypothetical sample, compare whether the retrieved passages contain the contract number, quantity, corresponding export order number and validity period together, and observe whether these fields appear inside one citation block. That outcome is treated as complete segmentation.
- Input the draft sentence "This supplier record can be used for input aggregation in the current refund declaration", compare whether the returned evidence contains the supplier taxpayer identification number and an invoice number, and observe that the returned passages do not state either item. Judge this as unsupported and retain the evidence gap.
- Input purchase contract number CG-2025-0417 as a full-text query, compare whether the result hits the passage containing that number, and observe whether the number is hit without passages from other contract numbers being mixed in. If the return is empty, record the empty result and check the index status rather than inferring a rule-based empty return. Libraries built on different vector models cannot be associated together; check the library identifier first, exclude those that do not match, and record the actual observation.