What the documents look like
Platform refund and return handling runs continuously per order, rather than being tied to a single live-stream commitment. The materials for one case arrive from separate systems: the refund application from the after-sales back office, chat records from customer service, receiving and inspection records from the warehouse, and refund records from payment. Each order produces its own set of documents, and daily operations add more.
For cross-border cases, one order can carry a Chinese-language refund application, a foreign-language return shipping label, and multilingual chat records at the same time. Retrieval therefore has to work across languages. The document parsing and chunking step turns these materials into data blocks that vector indexes and full-text indexes can reference, while preserving the document identity and version needed to trace a block back to its source. Responsibility decisions, return-chasing, and amount approval depend on the subject, object, status, and validity scope recorded in the material. If chunking separates a condition from its conclusion, the retrieved result cannot support the later manual review.
Constraints at the document chunking step
The smallest search unit in FastGPT is the library. A collection only classifies data and does not affect search results. One data group can map to multiple vectors, and when several vectors from the same data group are recalled, the vector score takes the highest value. Semantic retrieval supports cross-language understanding, so a Chinese question can match an English knowledge point. Full-text retrieval suits exact queries such as numbers, models, and proper nouns. Hybrid retrieval merges semantic and full-text results through RRF, and a ReRank model produces a 0-1 score used for filtering. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. The reference limit is controlled by tokens.
Retrieval and form collection do not constitute authorized execution. Approval, database writes, and external synchronization each need their own downstream implementation. A pending request is not an approval. An offered option is not execution.
Hypothetical exercise sample: the source material reads, "The refund application for order A-1042 was submitted by the buyer on 12 March 2025. Return waybill SF-7781 shows delivery signed on 15 March. The inspection photo is annotated as outer packaging damaged. The refund approval column remains pending review." The draft sentence to be judged reads, "The refund for this order has completed approval and has been refunded." The comparison object is the refund approval status field, the period is 12 March to 15 March 2025, and the authorization boundary is limited to an after-sales specialist checking within the target library. No actual refund execution is involved.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Material contains both waybill numbers and natural language; compare recall coverage against semantic retrieval alone |
| Minimum relevance | 0.3 | Hypothetical starting point; compare retained result count against 0.2 and 0.4 |
| Reference limit | 2000 tokens | Hypothetical starting point; compare reference completeness against 1500 and 3000 |
| Question optimization | Enabled | Compare retrieval results in follow-up question scenarios against the disabled state |
| Result reranking | Enabled | Compare ordering changes against the disabled state when candidates are numerous |
| Vector model | Consistent with the target library | Libraries with different vector models cannot be associated together; verify identifiers first and exclude non-matching ones |
Three common errors
- Treating a successful upload as an updated index. After material changes, the actual training and index status must be checked. The upload action and the index taking effect are two separate facts.
- Treating pending review as approved. In the sample, the refund approval column is pending review while the draft sentence states approval is complete and the refund has been issued. These are mutually exclusive values in the same field, so the draft sentence is unsupported.
- Treating a similarity setting as a permission filter. The minimum relevance value only filters low-relevance results. It does not exclude unrelated business records or unauthorized records. Retrieval itself is not authorization.
How to verify the setup
- Input the sample source material and the draft sentence, then compare the refund approval field value. If the field is explicitly pending review while the draft sentence claims completed approval, judge the draft as unsupported and record the evidence gap.
- Run a full-text search using waybill SF-7781, then compare whether the returned content includes the data block corresponding to that waybill. If no hit occurs, check whether the data has completed training and indexing. Do not assert a guaranteed first-hit result or complete recall.
- Search English return label content using a Chinese question, then compare whether semantic retrieval results cover the corresponding data block. If coverage is absent, record the actual observation without inferring cross-model association results. For empty-result handling, use a clearly defined paper input [] to check the downstream design, and do not claim that a fictional rule search necessarily returns [].