What the documents look like
For product questions in Certificates of origin, 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
For product questions within a certificate-of-origin application, use the proposed invoice’s selected coffee beans and one thousand kilograms as a name-and-quantity reading example. Mark the corresponding source wording and inspect whether candidates separate the product from its unit, followed by reranking with the complete question. Name and quantity do not supply origin-determination evidence. Retrieving an invoice cannot establish application eligibility; the reading record should retain each value and its source location. Distinguish a product description from evidence establishing its source. The coffee-bean name does not state a production location. A follow-up about origin needs the relevant material and a separate check, while the present answer stays within the supplied name-and-quantity evidence.
Configuration and implementation checks
| Item or implementation check | Input or applicable condition | Traceable check |
|---|---|---|
| Fused candidates | Coffee-bean name and one thousand kilograms | A product description does not supply production-location evidence. |
| Complete question | Total net weights of nine hundred fifty and nine hundred eighty kilograms | Check item, unit and version before comparison. |
| Minimum relevance | Inspect its effect during semantic retrieval or reranking | Relevance does not decide which net weight to adopt. |
| Citation token limit | Weight values and their total-net-weight scope | Avoid retaining numbers without their measurement meaning. |
| Missing-source retest | No shelf-life or storage-condition evidence | Obtain corresponding sources without inventing storage advice. |
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 proposed invoice and packing list give total net weights of nine hundred fifty and nine hundred eighty kilograms respectively. Preserve both values and file identities. Check product, unit and version comparability before inspecting whether citations retain “total net weight” rather than another weight scope. Two candidates appearing together do not explain the difference, and a higher rank cannot decide which weight to adopt. Continue the check against the corresponding original records.
How to verify the setup
The invoice attachment supplies neither shelf life nor storage conditions. Record both as material still required. Inspect whether nearby candidates are generic descriptions for another product and whether output invents dates or storage instructions. Once corresponding evidence is obtained, repeat the original question and trace its entry into candidates and citations. Check reranking and minimum relevance under their own conditions; they do not determine storage suitability or qualification for a certificate of origin.
| Material context | Scope |
|---|---|
| Document types | Certificate of Origin Application, Commercial Invoice, Packing List and Production Cost Data |
| Candidate sources | China International Trade Single Window, CCPIT System, Enterprise Production Records and Customer Customs Clearance Requirements |
| Technical baseline | FastGPT v4.17.2 |