What the documents look like
Producer export tax refunds combine the offset of domestic sales tax with the refund calculation. That combination shapes how product records and selection questions must be organized. A single export transaction carries a product name, quantity, free-on-board value and trade mode, and those values appear across several documents that have to point back to each other.
The working documents are the refund declaration detail schedule, export contracts, the sales detail ledger and export customs declarations. The declaration schedule is organized by line item, contracts by clause, the sales ledger by entry, and customs declarations by header plus commodity items. Document parsing and chunking breaks these materials into retrievable fragments that still allow a question to cite the original text.
Materials update monthly. Uploading a file and completing index training are two separate states, so after an update the retrieval result needs to be checked again against the current materials.
Constraints at the document chunking step
Chunking has to preserve the structure of each document type. If a condition, a unit of measure or the amount range it constrains is split away from the main content, the retrieved fragment cannot be judged on its own.
Original data and indexes are separate. One set of original data can map to several vector indexes. During retrieval the vectors are recalled first, then the vector identifier is used to find the original data in MongoDB, and records belonging to the same original data are merged with the highest score kept.
Hypothetical paper exercise. The contract text reads: "Export contract HX-2024-071 covers self-produced steel structure components, quantity 120 tonnes, transaction term FOB, contract signing date 3 June 2024." The matching sales ledger entry reads: "June 2024 export of steel structure components, 120 tonnes, total free-on-board value USD 960,000." The draft sentence to be judged is: "This batch of steel structure components has obtained refund approval."
In this exercise the contract and the ledger entry share the same subject, the same period and the same product, and the quantities agree. The draft sentence concerns refund approval status. Neither material states an approval outcome, so the sentence is not supported.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid search | Product names suit semantic recall, contract and declaration numbers suit full-text; compare number hit rate against a single mode |
| Reference limit | 3000 tokens as a hypothetical starting point | Declaration line items run long; compare whether fragments stay complete at 1500 and 3000 |
| Minimum relevance | 0.4 as a hypothetical starting point | Compare whether irrelevant fragments decrease at 0.2 and 0.4 |
| Question optimization | Enabled for follow-up turns | Compare whether "this batch" is completed to a specific contract number with and without it |
| Result reranking | Enabled when candidates are numerous | Compare rerank scores and vector scores for the same fragment |
| Chunking check | Split by line item and by clause | Compare whether each fragment can independently point back to the contract number and product name |
These values are hypothetical starting points for this exercise and define a comparison method. They are not official defaults.
Three common errors
- Treating a successful upload as an updated retrieval index. Input: a newly uploaded export contract. Compare the upload time against the training completion state, and observe whether retrieval still returns the previous version of the fragment. Without evidence of completed training, that state remains unverified.
- Treating a retrieval hit as business authorization. Input: the draft sentence "This batch of steel structure components has obtained refund approval." Compare the contract and the sales ledger entry for a stated approval conclusion, and observe that neither material carries an approval field, so the sentence is not supported.
- Associating libraries that use different vector models. Input: the vector model identifiers of two libraries. Compare whether the identifiers match, and observe that different identifiers rule out joint association. Cross-model retrieval is not executed.
How to verify the setup
- Input the contract text from the hypothetical exercise together with the draft sentence. Compare whether the retrieved fragment contains contract number HX-2024-071 and the product name, and observe whether the fragment can independently support a judgment about quantity and transaction term. If the fragment lacks the product name, the chunking is incomplete.
- Input the sales ledger entry text. Compare whether the retrieval result merges with the contract fragment into the same original data, and observe whether the quantity of 120 tonnes agrees across both materials. If only one is hit, record the recall gap.
- Input an explicit paper empty result []. Check the downstream question-answering design for handling content with no references, and observe whether it indicates that no material was retrieved. This check does not claim that a retrieval against invented rules must return an empty result.
Authorization boundary: retrieval and forms collect and return materials only. Approval, refund declaration, database writes and external synchronization each require their own downstream implementation. A pending review state does not show that an operation permission is absent, and signing, reaching an agreement and contract effectiveness are three distinct facts.