What the documents look like
Export tax rebate filing happens after goods leave the country, when tax vouchers are matched against export records. The main data sources are the electronic tax bureau, the export tax rebate filing system, customs electronic messages and the financial ledger. Daily documents include the export tax exemption and rebate filing form, the export customs declaration, purchase invoices and foreign exchange receipt vouchers. These materials frequently contain Chinese, English and overseas document formats, and declaration clerks alongside tax accountants need to compare product names, quantities, amounts and dates within one batch. Materials update monthly or per batch, and the quality of parsing and chunking directly affects the later work of matching invoices to customs declarations.
FastGPT builds its knowledge base on embedding vectors, and semantic retrieval depends on vector distance. The official documentation states that semantic retrieval supports cross-language understanding, for example matching a Chinese question to an English knowledge point, while precision is affected by model training quality, keywords and sentence completeness. A single data group can map to multiple vectors, and one original data record can carry several indexes, so a hit on any one vector recalls that data group. Chunking therefore needs to preserve the document identity and version that let a fragment point back to its source, and to keep original data distinct from retrieval indexes. If a condition statement is split away from the text it constrains, a retrieved fragment may lack its subject or its effective scope.
Constraints at the document chunking step
Multilingual and cross-border materials mean that one filing batch can contain files in different languages and different layouts. The official documentation notes that semantic retrieval suits natural-language questions and semantically similar content, while full-text retrieval suits identifiers, model numbers, proper nouns and error codes; hybrid retrieval combines both and merges results through RRF. For a rebate filing batch, customs declaration numbers and invoice numbers are exact identifiers, while product descriptions and matching questions are natural language, so the chunking step has to serve both retrieval paths without assuming one of them.
The reference limit is expressed in tokens: each search cites at most a set number of tokens of content. The documentation explains that token-based control was chosen over top k because chunk lengths vary greatly in mixed knowledge bases. Minimum relevance is a 0-1 value that filters low-relevance results and applies only under semantic retrieval or when result re-ranking is used. Neither setting performs entity isolation or authorization filtering. Retrieval returns candidate content; it does not verify a business fact, approve a filing, write to a database or synchronize with an external system.
Question optimization performs coreference resolution and question expansion before retrieval, at the cost of an extra model call. Re-ranking uses a ReRank model to reorder candidates and produces a 0-1 score that is generally more precise than a vector score. Image retrieval requires a corresponding image input and model capability, and re-ranking requires a model and must be enabled by configuration. The overall retrieval flow lists optional paths; not every request executes all of them.
Hypothetical exercise sample, clearly labelled as invented for this page: an export enterprise clerk prepares one batch of rebate materials. The source text available for checking reads: "Export customs declaration number HYP-2024-001, export date 15 March 2024, product name stainless steel kitchenware, quantity 1200 pieces. Purchase invoice number FP-2024-008, invoice date 20 March 2024, corresponding product name stainless steel kitchenware, quantity 1200 pieces." The draft sentence to be assessed reads: "The purchase invoice and export customs declaration for this batch of stainless steel kitchenware have completed matching." The draft sentence states only that two materials exist; it does not describe a matching operation or any authorization state. No matching action, approval, filing or system update is claimed to have occurred.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Identifiers and product names need exact hits, natural-language questions need semantic recall; compare recall differences between semantic retrieval and full-text retrieval |
| Reference limit | Hypothetical starting point of 3000 tokens | The official documentation controls citation volume in tokens; compare whether returned fragments cover the key fields of the customs declaration and the invoice at different values |
| Minimum relevance | Hypothetical starting point of 0.4 | The documentation states this value applies under semantic retrieval or re-ranking; compare candidates before and after filtering for irrelevant fragments |
| Question optimization | Enabled for short questions | The documentation describes coreference resolution and question expansion; compare recall completeness in follow-up questions with the setting on and off |
| Result re-ranking | Enabled when candidates are numerous | The documentation states re-ranking scores are 0-1 and more precise; compare ordering stability before and after re-ranking |
| Full-text retrieval | Used for identifier queries | The documentation states full-text retrieval suits identifiers, model numbers and proper nouns; compare hit results for identifier queries |
Three common errors
- Treating a successful upload as an updated retrieval index. Upload means the file entered the system; training and index state require separate checking. Compare the upload record against the fragments actually retrievable, and record a gap where the fragments are absent.
- Treating a similarity score as a permission filter. Minimum relevance is a filtering threshold and does not perform entity isolation or authorization judgment. Compare whether records belonging to different subjects are recalled together because their scores are close.
- Treating a pending review request as approved. When a draft sentence or handover request sits in a pending state, compare its status field against approved records and confirm whether execution conditions exist. A pending request is not approval.
How to verify the setup
- Input the hypothetical source text and the draft sentence given above, then compare whether the draft sentence exceeds what the source text states. The source lists two materials with matching product names and quantities; it does not state that matching was completed. A draft sentence claiming completed matching is unsupported. The expected observation is that the comparison identifies the unsupported claim; this page does not claim the test was run.
- Input the customs declaration number HYP-2024-001 as a full-text query, then compare whether the returned fragment contains that number and the corresponding product name. If the fragment lacks the number or the product name, record it as a retrieval gap. The expected observation is a fragment containing both; a missing field is a failure branch, and no guarantee is made that the first result is a hit.
- Use a clearly written paper input of [] to check empty-result handling downstream. Compare whether the downstream design treats it as an empty result. The documentation does not describe a rule that a fictional query must return an empty array, so this check records the downstream behavior rather than asserting a system outcome.
Where evidence is absent, it remains absent. The source text does not record an approval, a completed matching operation, a database write or an external synchronization, and none of these are asserted here.