What the documents look like
Producer export tax processes combine domestic sales tax offsets with export refund calculations. Product information and selection questions draw on four document types: the exemption, offset and refund declaration detail table, export contracts, sales detail ledgers, and export customs declarations. These materials are collected monthly. The same product appears under different measurement bases across them. A customs declaration records the free-on-board value and the commodity code. A sales detail ledger records revenue and quantity. A declaration detail table records the exemption, offset and refund amount. Retrieval must handle these cross-material differences in basis.
Selection questions are often short. A question may give only a product name and ask which document applies. The question itself lacks a subject and a period, which directly affects recall quality.
The following paper exercise sample is hypothetical and does not represent any real record.
Material one (hypothetical export contract excerpt): "Contract number H-2024-031, seller is a producer export enterprise, product is self-produced A-type motor, quantity 500 units, trade term is free on board, signing date 10 March 2024, both signature fields completed."
Material two (hypothetical sales detail ledger excerpt): "March 2024, A-type motor, outbound quantity 500 units, revenue recorded at free on board, corresponding contract number H-2024-031."
Proposed sentence for review: "The 500 A-type motors under contract H-2024-031 have completed export customs declaration and are included in the current period exemption, offset and refund declaration."
Assessment points: Material one states the signing and quantity. Material two states the outbound movement and recording basis. Neither material states the customs declaration completion status, and neither states that the batch has been included in a declaration. Under the same entity, same period and same product, the customs declaration and declaration inclusion facts are not supported and should be recorded as pending verification, not as a contradiction.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Check the vector model identifier first, exclude those that do not match, and record the actual observation. Do not run cross-model retrieval to compare missed recalls.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process, not a switch that can be turned on. The minimum relevance value takes effect only in semantic retrieval or when result reordering is used. It is a filtering threshold and does not perform permission or entity isolation. The reference upper limit is measured in tokens, so materials of different lengths have different completeness under the same limit. Question optimization performs coreference resolution and question expansion and adds one model call. Image retrieval requires image input and the corresponding model capability. Reordering requires a model and must be enabled by configuration. Optional paths shown in the overall document flow diagram do not mean every request executes all of them. The node output is an array whose length can be zero. When no content is retrieved, the output path still completes.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for selection and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the exemption, offset and refund material base with a matching vector model identifier | Check the identifier first, then compare against bases with non-matching identifiers and exclude them |
| Search mode | Start with full-text retrieval for commodity code questions | Compare against semantic retrieval results and observe differences in code matching |
| Reference upper limit | Hypothetical starting point of a token amount covering one detail table fragment | Compare against the completeness of a longer contract fragment, then adjust |
| Minimum relevance | Mark relevant and irrelevant samples first, then set a threshold | Compare against the marked results and observe whether irrelevant fragments are filtered out |
| Question optimization | Enable only for multi-turn follow-ups and short questions | Compare against recall content when disabled and check whether the subject is altered incorrectly |
| Result reordering | Enable when candidates are numerous and a reordering model is available | Compare against the order without reordering and observe changes in relevance sequence |
These values are hypothetical starting points for this exercise and are not official default values.
Three common errors
- Treating the minimum relevance value as a permission or entity filter. Erroneous input: using a relevance threshold to restrict materials from different enterprises. Comparison object: the node documentation states the value takes effect only in semantic retrieval or result reordering. Observable result: threshold adjustments change only the selection of low-relevance content, not the accessible range.
- Associating knowledge bases with different vector models together. Erroneous input: adding two bases with non-matching identifiers to the same node. Comparison object: the documentation requires associating knowledge bases with the same vector model. Observable result: after checking identifiers, exclude those that do not match. Do not infer runtime behaviour from a combined association.
- Treating a retrieval hit as completed business execution. Erroneous input: using a retrieved contract fragment to conclude that customs declaration is complete or that declaration inclusion has occurred. Comparison object: the sample materials state only the signing, quantity and recording. Observable result: customs declaration and declaration inclusion status are recorded as pending verification, not as completed.
How to verify the setup
- Verification one: Input the original text of material one and material two above, plus the proposed sentence for review. Compare the phrase "have completed export customs declaration and are included in the current period exemption, offset and refund declaration" against what the two materials actually state. Satisfied condition: when the materials do not state customs declaration and declaration inclusion status, the assessment result is recorded as pending verification with the evidence gap noted.
- Verification two: Input a short commodity code question and run it under full-text retrieval and semantic retrieval separately. Compare whether the reference content returned by each path contains the fragment corresponding to the code. Satisfied condition: record the actually observed matches and non-matches. Do not claim a first-position match or full recall.
- Verification three: Input a paper query with no corresponding material and check the node output. Compare the output array length against the downstream processing design. Satisfied condition: the output can be an array of length zero and the downstream path still completes. Do not claim that this query returns an empty result.
During operational handover, retrieval results and form entries are inputs pending verification. Retrieval does not constitute authorization. A form collecting values does not constitute business verification, approval, database writing or external synchronization. Those actions require their own downstream implementation.