What the documents look like
For product questions in General export declarations, this page examines FastGPT hybrid retrieval. Compare the material received with the sources used in the answer. The examples are hypothetical.
Constraints at the hybrid retrieval step
The export-name example provides “electronic components” in a proposed commercial-invoice attachment, while the question asks for comparison with a purchase order. First establish whether that order was actually supplied. An invoice passage matching the question cannot replace the other side. When both sources exist, rerank with the complete name-comparison question and retain the file identity of each candidate. Retrieving one invoice supports checking its stated name, without completing the requested comparison. The purchase order is another source required for this comparison, not a presumed knowledge-base field. Once obtained, verify that its actual wording enters candidates. Merely mentioning an order in the question does not provide its contents to retrieval.
Configuration and implementation checks
| Item or implementation check | Input or applicable condition | Traceable check |
|---|---|---|
| Fused candidates | Invoice components description and the purchase order still required | Do not claim a completed comparison without the order. |
| Complete question | Whether the product name corresponds to the original order | Check whether order text actually reaches reranking input. |
| Minimum relevance | Inspect the parameter under semantic or reranking conditions | An empty candidate set also requires a source-scope check. |
| Citation token limit | Two hypothetical price descriptions in invoice and contract | Keep both sources; ranking does not set a price. |
| Missing-source retest | Invoice lacks detailed technical specifications | Trace model requirements after adding the specification. |
Documented conditions
| Documented item / implementation check | Condition |
|---|---|
Hybrid retrieval | Vector and full-text results use RRF fusion. |
| Reranking | Reranking reorders candidates using the full question. |
Minimum relevance | Filtering applies to semantic search or reranking. |
Citation limit | Citation content is limited by tokens. |
Question optimization | Compare subjects and restrictions after optimization. |
Three common errors
The total-price example uses “price category A” and “price category B” as hypothetical wording, without an actual quotation. Preserve the statements from both invoice and contract, checking whether moving one passage forward causes the other to be omitted. Comparability still requires the same transaction, version and calculation scope. Neither RRF fusion nor later reranking makes a pricing decision; the returned wording needs to remain traceable for review by the responsible business party.
How to verify the setup
The existing invoice attachment does not contain detailed technical specifications, leaving model requirements without a source to match. Inspect whether apparently relevant candidates merely repeat the product name, then obtain the required specification and repeat the question. Keep evidence of the new source entering candidates and citations, including any long condition removed by the token limit. Check filtering under semantic or reranking conditions, without treating an invoice score as proof that a product meets the requested model specification.
| Material context | Scope |
|---|---|
| Document types | Export Declaration Form, Commercial Invoice, Packing List and Contract |
| Candidate sources | China International Trade Single Window, Customs Code Database, Freight Forwarder Bill of Lading and Enterprise ERP |
| Technical baseline | FastGPT v4.17.2 |