What the documents look like
TikTok Shop US store operations differ from shelf-based marketplace work because local return windows, return grading and overseas logistics shape what staff must look up. In store service and internal training, knowledge base retrieval handles four document types: product listings, return and refund records, Return Grading evidence, and settlement bills. These materials update daily or with each order. Retrieval requests typically come from after-sales specialists working on automatic returns, uploading image evidence, or checking settlement deductions.
The stated daily and per-order frequency describes how the source material itself is refreshed. It does not imply batch synchronization, full synchronization, real-time triggers, or transaction-triggered mechanisms.
Constraints at the retrieval step
The 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. Before running a search, staff check the vector model identifier, exclude non-matching bases, and record what was actually observed. A cross-model retrieval is not executed to compare missed recalls.
Reference content is output as an array and its length can be zero. When nothing is found, the output path still completes.
Search mode is compared across semantic, full-text and hybrid retrieval for the test question. No single mode is assumed to suit all material. The minimum relevance score is a filtering threshold and only takes effect under semantic retrieval or when result reranking is used. It does not perform permission or entity isolation. RRF is the fusion process for hybrid results and is not an on/off switch. Image retrieval requires corresponding image input and model capability. Reranking requires a rerank model and must be enabled by configuration. Optional paths shown in the overall flow diagram do not mean every request executes all of them.
Hypothetical drill sample: the material to be checked reads, "Store service record: object is US store after-sales ticket A-001, stage is return evidence pending upload, date range is 1 March 2025 to 7 March 2025, authorization boundary is view by the after-sales specialist for this ticket only. Return Grading evidence field states 'image evidence not uploaded'." The draft sentence to be judged is: "Ticket A-001's Return Grading evidence has been uploaded and can proceed to refund verification."
Configuration and implementation checks
The configuration below is tied to the sample above and lists only officially documented settings or implementation checks relevant to this feature and the primary task. Numeric values are hypothetical starting points for this trial, with the reason for choosing them and the comparison object stated.
| Configuration item | Hypothetical trial starting point | Check and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only bases that use the same vector model as the after-sales training material | Check the vector model identifier, exclude bases on different models, record the actual association result |
| Search mode | Run the sample question once each under semantic, full-text and hybrid retrieval | Compare whether the three return paths include the A-001 material; do not assume any mode necessarily hits |
| Reference limit | Measured in tokens; start with a value large enough to hold the full sentence of the sample material | Test with actual material of different lengths to see whether content stays complete, then adjust the limit |
| Minimum relevance score | Annotate relevant and irrelevant segments in the sample first, then take an initial threshold | Only effective under semantic retrieval or result reranking; compare whether relevant material survives filtering |
| Question optimization | Enable for follow-up questions containing references; disable for already explicit questions | Compare the original question with the optimized question, check whether the subject was altered incorrectly |
| Result reranking | Execute when a rerank model is available and enabled by configuration | Compare the rerank score with the vector score, check changes in candidate ordering |
Three common errors
- Associating knowledge bases built on different vector models. Input: two bases with different vector model identifiers. Compare whether the identifiers match. The observed result should be that non-matching bases are excluded and the actual observation is recorded; cross-model retrieval is not executed.
- Treating the minimum relevance score as a substitute for permission checks. Input: sample material whose authorization boundary is view by the after-sales specialist for this ticket only. Compare the retrieval return with the authorization scope. The observed result should be that retrieval does not perform permission isolation and authorization still requires downstream implementation.
- Reading a pending-upload state as already uploaded. Input: the material's original text "image evidence not uploaded" and the draft sentence "evidence has been uploaded". Compare the same field of the same ticket. The observed result should be that the two values are mutually exclusive and the draft sentence is unsupported.
How to verify the setup
- Input the sample material's original text and the draft sentence "Ticket A-001's Return Grading evidence has been uploaded and can proceed to refund verification". Compare the value of the same field of the same ticket. The observed result should be that the material states not uploaded while the draft sentence states uploaded; the draft sentence is judged unsupported and the evidence gap is retained.
- Input the sample material's original text and a handover request "Please transfer ticket A-001 to refund verification". Compare the authorization boundary field with the request object. The observed result should be that authorization is limited to viewing by the after-sales specialist, and transferring to refund verification exceeds that boundary; a separate system implementation and corresponding authorization are required. Retrieval or a form is not treated as authorization execution.
- Input an explicit paper-form empty result []. Compare the downstream path design. The observed result should be that when the reference content length is zero the output path still completes, and downstream handles it as an empty array. No claim is made that a fictional rule retrieval necessarily returns [].