What the documents look like
For product questions in Refund-only disputes, 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 proposed refund-only application names product A and gives “description mismatch” as the reason. Product identity and the stated reason are different information. A product question should reach the name field; a question about the complaint needs the reason’s original wording in candidates. Preserve that distinction in the full reranking question, so a nearby appeal summary does not turn an applicant’s statement into an established defect. This example contains self-drafted application text, without a dispute decision. Preserve the source role of an application reason in the wording, instead of abbreviating it into an established fact. Reading a product name is also separate from a later refund or appeal decision; neither follows from candidate order alone.
Configuration and implementation checks
| Item or implementation check | Input or applicable condition | Traceable check |
|---|---|---|
| Fused candidates | Product A and the stated description-mismatch reason | Inspect identity separately from the applicant’s statement. |
| Complete question | Relationship between application product B and appeal product C | Related dispute terms do not merge identities. |
| Minimum relevance | Inspect filtering under semantic or reranking conditions | A retained candidate is not a dispute determination. |
| Citation token limit | Product name and the source role of the stated reason | Do not convert an allegation into an established defect. |
| Missing-source retest | No model or specifications beyond the name | Obtain corresponding material for the selection question. |
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
Product B in the application and product C in the appeal require checking whether both documents concern the same matter. Candidates and citations need to retain their separate names. Similarity to “product dispute” does not make both passages one object. Inspect recall completeness first, then whether reranking conceals the other name. Without an order or other linking evidence, leave the identity unresolved rather than deciding which material to accept from its rank.
How to verify the setup
A name without model or specification information leaves the selection question without sufficient evidence. Inspect whether the final description borrows a model from another case or expands a refund reason into specific performance. Once corresponding specifications are supplied, repeat the question and trace new citations, recording candidate coverage separately from ordering changes. Check the conditions for minimum relevance; a highly scored application passage does not establish complete product-selection evidence.
| Material context | Scope |
|---|---|
| Document types | Refund-Only Application, Proof of Shipment, Proof of Receipt and Appeal Materials |
| Candidate sources | Platform After-Sales Backend, Carrier Tracking, Customer Service Chat and Warehouse Outbound Records |
| Technical baseline | FastGPT v4.17.2 |