What the documents look like
Operating review work for a TikTok Shop US store draws on four document families: product listings, return and refund records, Return Grading evidence, and settlement statements. These materials sit inside local return windows, return grading practice, and overseas logistics constraints, which shape what a review can and cannot confirm from a single record.
Data arrives from TikTok Shop Seller Center, the order center, carrier tracking, and advertising reports, with updates at daily and per-order frequency. In daily exception review, the slowest steps are handling automatic returns, uploading image evidence, and checking settlement deductions. Retrieval therefore has to serve two kinds of material at once: text fields such as order numbers, return reasons, and deduction status, and image evidence such as Return Grading photos.
A retrieval result is a reference. It shows what content matched a query. It does not show that a refund was executed, that a deduction was settled, or that a staff member holds authority to act.
Constraints at the retrieval step
In FastGPT, a knowledge base consists of libraries, collections, and data. The smallest search unit is the library, so a search runs across an entire library; collections only organize data and do not affect search results. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries built on different vector models cannot be associated together.
A retrieval pass can combine semantic search, full-text search, hybrid search, question optimization, and result reranking, then apply similarity filtering and a reference cap before producing reference content. Reference content is output as an array whose length can be zero, and the output path still completes when nothing is found. Semantic search relies on vector distance; full-text search relies on keyword hits; hybrid search merges the two paths through RRF. The minimum relevance value applies only in semantic search or when result reranking is used, and it is a filter threshold rather than a final ordering rule. Reranking produces a score from 0 to 1 for relevance. Retrieval carries no permission or entity isolation function.
The reference cap is measured in tokens, not in a count of items. Image retrieval depends on whether a vision model is configured or the vector model accepts image input, and on whether an image index was created when the material was stored.
Hypothetical paper exercise sample. A TikTok Shop US store after-sales specialist submits a handover request on 10 March 2025. The attached material text reads: "Order number US-20250305-0871. The buyer initiated a return on 8 March 2025 with the return reason entered as 'size mismatch.' Carrier tracking shows the parcel was signed for on 9 March 2025. Return Grading evidence photos have been uploaded to the order center. In the settlement statement, the deduction status for this order is 'pending verification.'" The draft sentence to be checked is: "This order has completed its refund and the deduction has been settled." The scope of checking is limited to this order's return and refund records and settlement statement fields between 5 March and 10 March 2025.
Configuration and implementation checks
| Configuration item | Assumed setting for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the return and refund records library and the settlement statement library that use the same vector model | The two libraries can be associated together only when their vector model identifiers match; check the model identifier library by library |
| Search mode | Hybrid search | Return reasons contain natural language while order numbers and deduction status contain exact fields, so both query types need coverage |
| Reference cap | Hypothetical trial starting value of 800 tokens for this exercise | Test content completeness with actual return records of differing length, and compare against smaller and larger values |
| Minimum relevance | Hypothetical trial starting value of 0.4 for this exercise | Label relevant and irrelevant samples first, then compare filtering results at this threshold |
| Question optimization | Enabled | The handover request contains references; compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable ReRank | When candidates are numerous, use the rerank score for filtering and compare against results with it disabled |
Three common errors
- Treating a retrieval hit as an authorization basis. Input "order number US-20250305-0871, deduction status pending verification" and conclude from the hit that the refund is complete. This treats a matched record as an executed business fact. The check is to compare the deduction status field for this order in the settlement statement; a pending verification status does not constitute settlement.
- Treating a pending review stage as absence of operating authority. The material shows the return request at a pending review stage, and the conclusion drawn is that the after-sales specialist lacks handling authority. This mixes a process stage with a permission fact. The check is to compare against authorization records; when no permission is stated, record the matter as unsupported.
- Forcing association of libraries built on different vector models. When the return and refund records library and the settlement statement library carry different vector model identifiers, an attempt is still made to associate them. This steps past the documented condition. The check is to verify identifiers first, exclude those that do not match, and record the actual observation.
How to verify the setup
- Input the order number US-20250305-0871 and the return reason "size mismatch" from the hypothetical sample, then compare whether the returned reference content includes this order's return and refund record. If the returned content contains this order number and the return reason field reads "size mismatch," retrieval covers that record. If the returned array is empty, record that this order was not recalled.
- Input the draft sentence "This order has completed its refund and the deduction has been settled," then compare the settlement statement fields in the returned content against the draft sentence. If the deduction status in the settlement statement reads "pending verification," the draft sentence does not agree with the material. If the returned content does not include the deduction status field, record the matter as unsupported.
- Input the material showing the return request at a pending review stage, then compare the retrieval results against authorization records. If the retrieval results contain no authorization field, record the matter as unsupported. If the returned content contains an authorization field with an approved status, check whether it concerns the same subject, the same period, and the same field.