What the documents look like
Operating reviews for a TikTok Shop US store sit inside 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. Data arrives from Seller Center, the order center, carrier tracking, and advertising reports, with daily and per-order update frequency. The citation tracing step has one job: check whether each statement in a report or review conclusion can be matched to specific source text, and record the material version and its valid scope.
Return grading evidence and settlement statements are separate documents with separate fields. A return grading status describes where a returned item stands in the grading process. A settlement statement describes amounts and platform actions such as automatic refunds. These two document types can share an order number while carrying different fields, so a conclusion that treats them as one combined fact needs a specific check against both originals.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model, and it outputs cited content as an array whose length may be zero. An empty array means the search path completed without returning citations; it does not mean the workflow raised an error and stopped. Hybrid search merges multiple recall paths through RRF, result reranking uses a ReRank model for a second ordering pass, and the minimum relevance value applies only under semantic search or result reranking. Retrieval and similarity scores do not perform permission or entity isolation, so every displayed citation must be checked against its source text one by one.
Hypothetical paper exercise. Material A is a March 2025 return record for a US store. Its source text reads: "Order number US-2025-0312-001; the buyer initiated a return on March 12, 2025; carrier tracking shows the parcel was picked up on March 15, 2025; return grading status is pending review." Material B is a settlement statement for the same period. Its source text reads: "Order number US-2025-0312-001; on March 20, 2025 the platform issued an automatic refund of USD 19.99." The draft sentence under review reads: "The order has completed the return and refund, and return grading has passed."
The check subject is the same order number, the period is March 2025, and the fields are return grading status and refund record. Material A shows pending review. Material B shows an automatic refund. The two materials carry different fields, and the draft sentence upgrades pending review to passed. That statement is unsupported by the supplied text. The refund record does not supply a grading outcome, and the grading status does not supply a completed refund conclusion.
Configuration and implementation checks
| Configuration item | Use on this page | How to check |
|---|---|---|
| Associated knowledge bases | Select knowledge bases that already contain return and refund records and settlement statements | Check the selected knowledge bases and material versions; only knowledge bases with the same vector model can be associated together |
| Search mode | Use full-text search or hybrid search for return order numbers and settlement numbers | Compare semantic search and full-text search recall on identifier-style queries |
| Minimum relevance | Filter low-relevance citations | Applies only under semantic search or result reranking; record the actual threshold used |
| Citation limit | Control the total tokens cited in one search | Set by tokens as documented, not by estimated item count |
| Question optimization | Complete references in multi-turn follow-up questions | Check whether the completed question matches the original question |
| Cited content output | Carry the search node array output downstream | Check whether the downstream answer corresponds to cited source text item by item |
For this hypothetical exercise the starting minimum relevance is set to 0.5. The reason is that this value sits in the middle of the 0-1 range, which makes it easier to observe how the number of citations changes before and after filtering, compared with the result when the value is not applied. This value is a hypothetical starting point for this exercise only and is not an official default.
Three common errors
- Merging the pending-review return grading status in Material A with the automatic refund record in Material B and concluding that the return and refund are complete. The input is the source text of both materials, the comparison is between the return grading status field and the refund record field, and the observable result is whether the draft sentence rewrites pending review as passed.
- Treating the citation array returned by search as an authorization credential, so that the presence of a citation is read as approval of an operation. The input is the search node output, the comparison object is the stated authorization boundary, and the observable result is whether retrieval results are conflated with approval status.
- Forcing an association between knowledge bases that use different vector models, or inferring missed recall or redundancy after such an association. The input is the knowledge base identifier and the vector model identifier, the comparison object is the same-vector-model condition, and the observable result is whether the mismatch is recorded and the knowledge base is excluded instead of being searched across models.
How to verify the setup
- Use hypothetical Material A and Material B. The input is the source text of both materials. Compare the field values in the draft sentence against the source text. The condition is met when the draft sentence does not state pending review as passed and the evidence gap is recorded.
- Use an explicit paper input of an empty array []. Check whether the downstream answer step handles the empty citation output as an empty result. The condition is met when the downstream step does not raise an error because of the empty array and does not fabricate cited content.
- Use two knowledge bases with different vector model identifiers. Check whether the association is blocked or requires excluding the mismatched knowledge base first. The condition is met when the mismatch is recorded and the knowledge base is excluded, without running a cross-model search to compare missed recall.
These checks concern the citation tracing task itself. Whether any operation, request, or handover is authorized depends on the stated authorization boundary of the relevant material, and retrieval output or a form entry does not by itself establish that boundary.