What the documents look like
After-sales work for a TikTok Shop US store is shaped by local return windows, return grading, and overseas logistics constraints. The working materials are product listings, return and refund records, return grading evidence, and settlement statements. These update daily and per order. Three retrieval problems follow from that material set.
Return window terms and grading conclusions sit in separate records, so a single order requires cross-checking the refund, the return fee, and the settlement deduction. Whether image evidence was indexed at ingestion determines whether it can be retrieved later. Settlement deduction lines must be matched item by item against return records. The knowledge base search node can reference one or more knowledge bases that use the same vector model. Search operates on the knowledge base as the smallest unit; a collection only classifies data and does not change retrieval results.
Constraints at the retrieval step
The constraints come from the evidential state of the materials themselves. Refund status, return grading conclusions, and settlement deduction amounts are distinct fields. A judgment requires the same subject, the same period, and the same field; only explicitly conflicting values within one field form a contradiction. Evidence that covers only a related stage is unsupported or pending verification. A pending review does not show that an operating permission is absent. Signing, reaching cooperation, and contract effectiveness are three separate facts. The same-vector-model requirement constrains only which knowledge bases can be referenced together; it does not extend to external systems.
Semantic retrieval and relevance scores order and filter results. They do not perform permission or entity isolation, and a retrieval hit is not a business authorization.
Hypothetical paper exercise sample. The source text reads: "Order US-2024-0871 was signed for on June 3, 2024, with the return window open until June 17, 2024; the return grading conclusion is resellable; the June 2024 settlement statement deducts a return handling fee of USD 4.20." The draft sentence to be judged reads: "The return window for this order is closed and the refund has been executed." Comparing the draft with the source: the source states only the signed-for date and the window end date, and carries no refund execution record. Both "is closed" and "has been executed" have no corresponding field, so they are unsupported and the evidence gap is recorded.
Configuration and implementation checks
The table below lists settings or check items tied to this sample. The values are starting points for this hypothetical exercise, used for comparison against another set of values, and are not official defaults.
| Setting or check item | Starting point for this exercise and reason | Comparison and observation method |
|---|---|---|
| Referenced knowledge bases | Include the bases holding return and refund records and return grading evidence, and confirm the vector model identifiers match | Compare base identifiers and vector model identifiers, exclude mismatches, record the actual referenceable scope |
| Search mode | Start with hybrid retrieval because the sample contains an order number and natural-language description | Run the same question through semantic retrieval and full-text retrieval, compare whether candidates cover the order number and window fields |
| Reference limit | Set a starting token amount covering one return record plus one settlement record, since the sample needs two fields side by side | Test with actual materials of different lengths, observe whether referenced content is complete, then adjust |
| Minimum relevance | Start at a middle value to observe filtering behavior | Label relevant and irrelevant samples first, then adjust the threshold and observe whether irrelevant passages are filtered |
| Question optimization | The sample question contains the reference "this order"; after enabling, check whether the subject is altered | Compare the original and optimized questions, confirm the order number is not replaced or dropped |
| Empty-result handling | Use a paper input of an empty reference array to check downstream design | Confirm the output path still runs when reference length is 0, without claiming that a rule search must return empty |
Three common errors
- Treating a retrieval hit as business authorization. Retrieving a return and refund record does not mean the refund was executed or the return window is closed. The check must locate a refund execution field or a window status field; when absent, record it as unsupported.
- Treating minimum relevance as a permission filter. Relevance is a 0-to-1 filter threshold that applies only in semantic retrieval or when result reranking is used. It cannot exclude unrelated businesses or unauthorized records. Compare candidate content for the same question under different thresholds and observe whether filtered items are low-relevance text.
- Forcing knowledge bases with different vector models into one reference set. Check each base's vector model identifier before referencing, exclude mismatches, and do not run cross-model retrieval to compare missed recalls. Record the observed referenceable scope and the reason a base cannot be referenced.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence, then compare "is closed" and "has been executed" in the draft against the signed-for date, window end date, grading conclusion, and deduction amount actually stated in the source. The condition is met when the draft sentence is marked unsupported and the check record retains the note that the refund execution field is missing.
- Input the same order number and run semantic retrieval and full-text retrieval separately, comparing whether the returned references cover the order number and the return window field. The condition is met when candidates from both modes can be checked and the differences are recorded, without claiming that one mode must match.
- Input an empty reference array and check whether the downstream path runs, comparing whether the output reference length is 0. The condition is met when the path still executes at length 0 and the empty result is not interpreted as a rule search necessarily returning empty.