What the documents look like
Producer export tax refund offsets domestic sales tax against export refunds in one combined calculation. The work of collecting self-produced export data, computing the offset-and-refund amount, and reconciling free-on-board price differences takes most of the time. The recurring materials are the offset-and-refund declaration detail table, export contracts, the sales detail ledger, and export customs declarations. Data comes from the electronic tax bureau, ERP sales records, customs electronic messages, and the financial ledger, and it updates monthly.
In an operations handover these materials serve two roles at once. They are retrieval targets, and they are written carriers of authorization status. The knowledge base search node returns reference content only, and the reference content array may have length zero, meaning the output path still runs even when nothing is found. Whether a material has been authorized, sits in pending review, or has been fully signed must be checked separately. Retrieval does not constitute authorization.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The vector retriever is the PostgresSQL PG Vector extension with an HNSW index. MongoDB stores original data in the dataset.datas table and records an indexes field. Knowledge bases built on different vector models cannot be associated together. When candidate bases appear, check the vector model identifier first, exclude those that do not match, and record what was observed. 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. Reranking uses a ReRank model to sort candidates a second time and produces a score between 0 and 1. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filtering threshold rather than a permission or entity isolation control. The reference limit is measured in tokens and does not use top k. Question optimization performs coreference resolution and question expansion and adds one model call. Image retrieval requires a corresponding image input and model capability, and reranking requires a model enabled by configuration. Optional paths do not mean every request executes all of them.
Hypothetical paper exercise sample: an export producer hands over a batch of October 2024 offset-and-refund materials. For checking, the export contract excerpt reads: "Party A is a domestic export producer, Party B is an overseas buyer; the subject matter is one batch of self-produced goods, and the free-on-board price follows the customs declaration; this contract takes effect from the date both parties sign it." The offset-and-refund declaration detail table excerpt reads: "Period October 2024, export sales amount is filled in according to the free-on-board price on the customs declaration." The draft sentence to be judged is: "The above export contract has taken effect, the corresponding offset-and-refund declaration detail table has completed declaration, and after handover the receiving person continues to process the refund." The handover request to be judged is: "Please have the receiving person directly execute the refund declaration operation based on the above two materials." Neither material states a signing date, a declaration submission status, or an operational authorization scope.
Configuration and implementation checks
The table below concerns the sample above and lists only settings or implementation checks that the official documents describe. The values are hypothetical starting points for this exercise, used to compare against another value, and are not official defaults.
| Configuration item | Hypothetical starting point and reason | Comparison method |
|---|---|---|
| Associated knowledge bases | Associate only bases sharing the vector model of the sample materials; check the identifier first | Compare the vector model identifier against each candidate base, exclude mismatches, and record |
| Search mode | The sample contains contract clauses and declaration fields, so start with hybrid retrieval | Compare reference content item by item against semantic retrieval alone and full-text retrieval alone |
| Reference limit | Start at a token amount that can hold the full contract effectiveness clause and the full declaration period wording | Compare against a smaller value and observe whether reference content is truncated |
| Minimum relevance | Start at a mid-range value inside the 0 to 1 interval | Compare against higher and lower values and observe whether irrelevant fragments increase or decrease |
| Question optimization | The sample question contains the reference "the above", so enable it first | Compare with it disabled and check whether the subject is altered incorrectly |
| Result reranking | Enable when candidates are numerous, and filter by its score | Compare candidate ordering with reranking disabled and check the score interval |
Three common errors
- Treating a retrieval hit as completed authorization. The sample contract text states only that it takes effect from the date both parties sign it, and no signing date is recorded. The draft sentence claiming it "has taken effect" is unsupported. A returned reference does not mean the signing fact has been verified.
- Writing a pending review status as a lack of operational permission. The sample does not state the declaration submission status, so the draft sentence claiming "has completed declaration" lacks a basis. A pending review does not show a lack of operational permission; both facts need separate checking.
- Forcing knowledge bases on different vector models into one association. The vector model identifier of the base holding the sample materials must be checked first, and mismatches are excluded. Do not infer from this that a joint association would produce missed recalls or redundant results.
How to verify the setup
- Input the sample export contract text and the draft sentence "the above export contract has taken effect". Compare whether the source text states a signing date and an effectiveness condition. Observable result: when the source text does not state a signing date, the "has taken effect" wording is judged unsupported, and the evidence gap is recorded.
- Input the sample offset-and-refund declaration detail table text and the draft sentence "has completed declaration". Compare whether the source text states a declaration submission status and a period field. Observable result: when the source text contains only "Period October 2024", the declaration completion status is judged as pending verification, and it must not be judged as a denial or a conflict.
- Input the handover request "please have the receiving person directly execute the refund declaration operation based on the above two materials". Compare whether the authorization scope the request relies on is stated in the materials. Observable result: when the materials do not state an operational authorization scope, the request sits in an unauthorized state, and neither retrieval nor a form constitutes an execution basis. Separately, use an explicit paper input of an empty reference array to check the downstream design, without claiming that retrieval on a fictional rule necessarily returns an empty array.