What the documents look like
Producer export tax processes combine the offset of domestic-sale tax amounts with the refund of export tax, so the calculation path differs from that used by trading companies. Contract and transaction material comparison for this category works across four document types at once: the exemption, offset and refund declaration detail schedule; the export contract; the sales subsidiary ledger; and the export customs declaration. These materials are produced at different stages, and their field definitions and issue dates do not align. Comparison work covers the aggregation of self-produced exports, the exemption-offset-refund calculation, and the handling of free-on-board price differences. If the parsed result of any material departs from its source text, the deviation carries into the tax amount calculation, which can lead to a reduced refund or a corrected tax filing. Materials update monthly, and the comparison scope normally falls within the same period.
Chunking for comparison must preserve document identity and version so that each chunk can state which material and which period it came from. Segmentation should follow the structure of each material; a condition, an explanation, and the content it constrains should not be split apart. If the trade term and free-on-board price basis in an export contract are placed in different chunks from the corresponding fields on the customs declaration, the constraining relationship becomes hard to reconstruct during comparison. Data and indexes are distinct: one group of source data can map to multiple indexes, the degree to which index content is condensed affects semantic recall precision, and the search scope is set by the whole knowledge base, with collections serving only as classification management. Semantic recall and relevance filtering act only on content similarity; they do not perform entity, period or authorization isolation, and retrieval itself does not constitute authorization. The search mode should follow the query type: identifiers such as declaration numbers rely on full-text or hybrid search, natural-language descriptions rely on semantic search, hybrid results are merged through RRF, and reranking requires a model and must be enabled according to settings. After material changes, the actual training and index state must be checked; a successful upload does not mean retrieval has been updated.
Constraints at the document chunking step
The primary task on this page is contract and transaction material comparison, and the constraints below apply to the parsing and chunking step that supports it. A chunk that cannot be traced back to a source document and version cannot be used to confirm which material and period a field belongs to. Conditions and the content they constrain belong in the same chunk, so that a trade term is not separated from the price basis it governs. Identifier queries and natural-language queries follow different recall paths, and the choice of path changes what a comparison can observe. Relevance thresholds operate on similarity scores and do not exclude unrelated businesses or unauthorized records. Where a field is absent from the source text, or where a period is not stated, the correct record is an evidence gap rather than a negative or conflicting finding.
Configuration and implementation checks
The configuration below concerns the hypothetical exercise sample used on this page. The numeric values are starting points for this exercise only, chosen so that observed return values can be compared against them; they are not official defaults.
Hypothetical exercise sample: the source material text reads, "Party A is a production enterprise; in June 2025 it exported a batch of self-produced goods, export contract number HT-2025-0601, trade term FOB, free-on-board price USD 120,000; the corresponding export customs declaration number is 202506010001, and the sales subsidiary ledger records June 2025." One draft sentence to be assessed reads, "HT-2025-0601 corresponds to customs declaration 202506010001, and the free-on-board price is settled on a cost-and-freight basis." The draft sentence takes a different value from the source text on the trade term field.
| Configuration item | Setting | Basis and check method |
|---|---|---|
| Knowledge base scope | Place the four material types in the same knowledge base | Search operates at knowledge base level and collections are classification only, so the sample materials must be comparable within one knowledge base |
| Search mode | Use full-text or hybrid search for identifier queries | Declaration numbers and contract numbers are exact queries, and semantic search hits identifiers unstably |
| Minimum relevance | Starting point for this exercise 0.3 | Compare against observed returned content to see whether relevant chunks are filtered out; this value takes effect only in semantic search or reranking |
| Citation limit | Starting point for this exercise 3000 tokens | Compare against the length of a single returned citation set to control truncation; top k is not used for this control |
| Chunking method | Segment by material structure and field boundaries | Keep conditions and the content they constrain in one chunk, and preserve document identity and version |
| Post-update check | Check training and index state after changes | A successful upload is not treated as evidence that retrieval has been updated |
Three common errors
- Placing the contract number and the customs declaration number in different chunks, so that their correspondence cannot be confirmed within one chunk during comparison. Check: input the sample contract number and declaration number, compare whether the returned chunk contains both identifiers, and treat the condition as met only when both appear in the same returned chunk.
- Using semantic search for an identifier query and relying on relevance filtering, then treating a missed identifier as proof that the material does not exist. Check: input declaration number 202506010001, compare the returns from full-text search and semantic search, and treat the condition as met only when full-text search hits the chunk containing that identifier.
- Treating a successful upload as retrieval already updated, without checking training and index state. Check: after changing a material, input a term that appeared in that material, compare returned content before and after the change, and treat the condition as met only when the return reflects the changed content.
How to verify the setup
- Verification one: input the sample contract number HT-2025-0601, compare whether the returned chunk contains the trade term FOB and the free-on-board price USD 120,000, and treat the condition as met only when both fields appear in the chunk. If the returned chunk contains only the contract number while the fields are absent, record this as incomplete chunking.
- Verification two: input the draft sentence "HT-2025-0601 corresponds to customs declaration 202506010001, and the free-on-board price is settled on a cost-and-freight basis," then compare the source trade term field with the value in the draft sentence. The two take mutually exclusive values on the same field, which is judged a contradiction. If the draft sentence states only that the contract number corresponds to the declaration number, that part is not supported by the source text and remains pending verification.
- Verification three: input an empty citation list [] and check how the downstream design handles an empty result, observing whether the return is empty and how the following process handles it. This result does not mean that any rule retrieval necessarily returns empty. Where a field is missing, a period is not stated, or only evidence from a related stage exists, record an evidence gap rather than a negative or conflicting finding.
Authorization status at handover remains a separate matter: which materials, operations or requests are in what authorization state cannot be inferred from retrieval or from a form. Missing evidence stays missing.