What the documents look like
TikTok Shop US operations records centre on product listings, return and refund records, Return Grading evidence, and settlement statements. These materials also carry the state of operational permissions and handover status. United States return windows, return grading, and overseas logistics constraints make return and refund records and Return Grading evidence the materials most likely to generate permission disputes during handover. Records update daily or per order, and handover checks must be completed within the scope of the specific document at hand.
A handover check needs to distinguish several states that appear in the same record set: a refund status field, an after-sales application status field, and the grading conclusion attached to the returned item. Retrieval can surface all three from the same order, and the retrieved text does not itself indicate which state governs whether an action may proceed.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The vector model identifier must be checked first, and non-matching bases excluded; cross-model retrieval must not be run to compare missed recalls.
Cited content is output as an array and its length can be zero, meaning the output path still runs even when no content is found. Semantic retrieval and the minimum relevance score affect content relevance only. They do not isolate permissions, entities, or authorization states, and a retrieval hit does not mean an operation has been authorized. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process, not a switch. Reranking requires a configured model and must be enabled according to settings.
Hypothetical paper exercise sample
Source text to be checked: "For the period 1 March 2025 to 31 March 2025, in the US store return and refund records, order US-2025-0312-018 has a Return Grading conclusion of 'resellable', image evidence has been uploaded, the refund amount is USD 48.00, and the refund status is 'refunded'. The after-sales operation permission record for this order shows: after-sales specialist A submitted the refund application on 15 March 2025, with status 'pending review'."
Draft sentence to be judged: "The refund for this order has been approved and executed by after-sales specialist A."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the return and refund records base already verified as using the same vector model | Check the vector model identifier first, exclude non-matching bases, then compare retrieved content |
| Search mode | Test semantic retrieval and full-text retrieval separately | Compare retrieved content per test question; do not assume one mode suits all materials |
| Citation limit | Hypothetical starting point of 2000 tokens for this exercise | Test content completeness with materials of different lengths, then compare against adjacent values |
| Minimum relevance | Hypothetical starting point of 0.4 for this exercise | Label relevant and irrelevant samples first, then compare the effect of threshold changes on results |
| Question optimization | Enable for short questions and multi-turn follow-ups | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable according to settings when candidates are numerous | Compare ordering before and after reranking and confirm the model is configured |
Three common errors
- Treating a retrieval hit as authorization. The input is the "pending review" refund application in the sample above. Judging from it that the refund has been approved is an error; the application status field and the approval completion record must be compared and they differ.
- Associating knowledge bases that use different vector models. The input is two bases with different vector model identifiers. Associating them directly and comparing retrieved results is an error; identifiers must be checked first and non-matching bases excluded.
- Treating RRF as a switch that can be turned on. The input is a hybrid retrieval configuration. Looking for an RRF switch in the configuration is an error; RRF is a fusion process and has no separate switch.
How to verify the setup
- Input the sample source text above and the draft sentence. Compare the refund status field "refunded" with the application status field "pending review" and observe whether the two are distinguished. If the draft sentence still states that approval and execution occurred, the condition is not met.
- Input two knowledge bases with different vector model identifiers. Compare the identifiers and observe whether the non-matching base is excluded. If they are still associated together, the condition is not met.
- Input an empty citation array []. Compare whether the downstream design accepts an empty result and observe whether the path runs through. If the downstream design does not define handling for an empty result, the condition is not met.