What the documents look like
TikTok Shop US operations differ from Douyin shelf operations because of local return windows, return grading, and overseas logistics constraints. The working materials for this category are product listings, return and refund records, return grading evidence, and settlement statements. These materials update daily or per order. For contract and transaction material comparison, the task is to check the return record, the grading evidence, and the settlement deduction for the same order, in the same period, under the same field. Retrieval only recalls candidate materials. Authorization status, refund execution, and appeal submission belong to downstream processes.
A hypothetical paper sample is used throughout this page. Return record text: "The buyer initiated a return on March 10, 2025, for product A Bluetooth earphones, quantity 2." Return grading evidence text: "Received March 14, 2025, appearance intact, function normal, resalable." Settlement statement text: "March 15, 2025, return processing fee of 12 USD deducted." One draft sentence to be judged: "The returned product for this order has been judged not resalable, and the platform should refund the full payment." The comparison object is the same order, the same period, and the same field: returned product condition.
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 using different vector models cannot be associated together. Check the vector model identifier first, exclude those that do not match, and do not run cross-model retrieval to compare missed recalls. Retrieval output is an array of cited content, and its length can be 0, meaning the downstream path still runs even when nothing is found.
Semantic retrieval, full-text retrieval, and hybrid retrieval are compared per test question. No single mode is assumed to suit all materials. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold, not a final ranking rule, and it does not perform permission or entity isolation. The citation limit uses tokens as its unit; test content completeness with actual materials of different lengths. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered incorrectly.
RRF is a fusion process for mixed results. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled according to settings. The overall retrieval flow lists optional paths; these do not all execute on every request.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the return materials base that uses the same vector model | Check each base's vector model identifier, exclude different models, and do not run cross-model retrieval |
| Search mode | Use full-text retrieval for identifier-type questions and hybrid retrieval for status descriptions | Observe recalled content per test question; do not assume one mode fits all |
| Citation limit | Start at 4000 tokens | 4000 tokens is a hypothetical trial starting point. Source length has not been measured. Measure the candidate excerpts with the selected model's tokenizer, compare completeness at 2000 and 6000 tokens, and adjust from those observations |
| Minimum relevance | Start at 0.3 | Label relevant and irrelevant samples first, then compare whether irrelevant fragments decrease across thresholds |
| Question optimization | Enable for multi-turn follow-up scenarios | Compare original and optimized questions; check whether the order number and product subject were altered |
| Result reranking | Enable when candidates are numerous | Observe ranking changes using the 0-1 rerank score and compare with results without reranking |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input an unrelated order query, compare whether returned content still includes other order records, and observe whether a low score is wrongly read as excluding unauthorized materials. A relevance score does not perform authorization judgment.
- Treating a pending review or a quotation validity period as already effective. Input the draft sentence "should refund the full payment," compare the status fields in the return record and settlement statement, and observe whether a pending-verification state is written as an approved refund. A pending review does not prove the absence of operating permission, and a quotation validity period does not indicate a supply term.
- Treating an empty result array as a failure. Input a paper empty array [], compare whether the downstream path still runs, and observe whether an empty citation is wrongly judged as a retrieval fault. A citation length of 0 is a documented output behavior.
How to verify the setup
- Input the return record text above and the draft sentence "The returned product for this order has been judged not resalable, and the platform should refund the full payment." Compare whether "resalable" in the grading evidence and "not resalable" in the draft sentence are mutually exclusive values under the same field. Only when the same order, same period, and same field clearly hold mutually exclusive values does the condition for a contradiction hold. In this sample the grading text and the draft sentence are opposite, so this is a contradiction.
- Input the settlement statement text "March 15, 2025, return processing fee of 12 USD deducted." Compare whether the statement field records a refund amount or refund status. Only when the statement records a refund field does the condition for supporting the refund claim hold. In this sample it is not recorded, so the claim is unsupported and the evidence gap must be recorded.
- Input a paper empty array [] and compare whether the downstream path continues to execute. Only when the downstream receives empty citations and still runs normally does the condition for empty-result handling hold. Whether retrieval itself returns an empty array must be based on actual observation; no claim is made that a fictional rule necessarily returns [].