What the documents look like
TikTok Shop US operations sit under local return windows, return grading, and overseas logistics constraints, so inventory procurement and supplier records are interwoven with return and refund records, Return Grading evidence, and settlement statements. The source systems named in the brief are TikTok Shop Seller Center, the order center, carrier tracking, and advertising reports, with daily and per-order updates. Retrieval questions in this task commonly involve supplier names, purchase batches, return attribution, and deduction reconciliation. When these fields cross over during retrieval, downstream appeal material can cite the wrong object.
A knowledge base search node can be associated with one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; the identifier is checked first and non-matching bases are excluded. Search modes are compared against test questions across semantic, full-text, and hybrid retrieval, and no single mode is presumed suitable for all material. Hybrid retrieval merges multiple result channels through RRF, which is a fusion process rather than a switch. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it is a filtering threshold that does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to external systems. The citation limit is measured in tokens, and content completeness is tested with material of differing lengths. Question optimization performs coreference resolution and question expansion, and the original question is compared with the optimized question to check whether the subject was altered. Cited content is output as an array whose length can be zero, and the downstream chain still runs when nothing is retrieved.
Hypothetical paper exercise: a paper record states that Supplier A shipped purchase batch P-031 to the US store in March 2025, quantity 200 units; Return Grading evidence for return order R-118 has been uploaded; settlement statement S-207 shows the deduction for this batch is pending verification. A proposed sentence reads: "Supplier A has confirmed responsibility for return R-118 and agreed to bear the deduction." The record shows only that evidence was uploaded and the deduction is pending verification. It does not state that the supplier confirmed responsibility or agreed to bear the deduction, so the proposed sentence is unsupported.
Constraints at the retrieval step
The search node accepts one or more knowledge bases sharing the same vector model. When a candidate base uses a different vector model, the identifier is checked and that base is excluded before any search is run. This page does not describe what happens if different-model bases are associated together, because the documented condition is that they cannot be associated.
Search mode selection is a comparison task. The same test question is run under semantic retrieval and under full-text retrieval, and the recalled content is compared. Number-like terms and natural-language phrasings can behave differently, and the comparison records which mode returned the target record rather than assuming one mode is universally better. Hybrid retrieval uses RRF to merge semantic and full-text results, and RRF is described as a merging formula, not a toggle. Minimum relevance applies only under semantic retrieval or result reranking; it filters low-relevance results and is not a permission or entity filter. The citation limit is expressed in tokens, not in a top-k item count, because chunk lengths differ widely in mixed knowledge bases. Question optimization runs a model call before retrieval to resolve references and expand the question, and it can increase latency, so its use is weighed against the specific conversation.
The downstream chain receives cited content as an array that may be empty. An empty array is a valid output, and the chain continues. Whether a downstream step treats the empty array as a miss, a fallback, or a stop is a separate implementation decision that this page does not assert.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the inventory procurement and supplier records base | Check each base's vector model identifier and exclude non-matching bases; then compare whether retrieved results come only from the associated base |
| Search mode | Run semantic retrieval and full-text retrieval separately | Use the same test question under both modes and compare recalled content, observing differences between number-like terms and natural-language phrasings |
| Citation limit | Starting point set to cover the token volume of one complete return record | Test with procurement and return material of differing lengths, compare whether content is truncated, and compare against the full source text |
| Minimum relevance | Starting point set to a middle value | Label relevant and irrelevant samples first, then adjust the threshold and compare whether the target record is still retained after filtering |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether supplier names or batch numbers were altered |
| Result reranking | Enabled when candidates are numerous | Compare ordering before and after reranking and observe the rerank score and the position of the target record |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input the same test question and compare whether returned content includes records for other suppliers. The relevance score does not perform isolation, and this check records what is observed rather than asserting that unrelated records are always excluded or always returned.
- Associating knowledge bases that use different vector models. Check the identifier, exclude non-matching bases, and record the actual observation. This page does not claim that such association produces missed recalls, redundant results, or unpredictable behavior, because the documented condition is that these bases cannot be associated.
- Writing a pending-verification state as a confirmed state. Compare the proposed sentence against the source record and check whether it adds a supplier confirmation of responsibility or an agreement to bear the deduction. Content not stated in the source is recorded as unsupported.
How to verify the setup
- Input the hypothetical source record and the test question "Which return order does P-031 correspond to?" Compare whether the returned citation contains R-118 with Supplier A as the subject. If a different batch or an empty array is returned, record it as a miss. First-position retrieval is not guaranteed, and all-recall is not guaranteed.
- Input the proposed sentence and the source record. Compare whether both share the same subject, the same period, and the same field. The source does not state that the supplier confirmed responsibility, so the proposed sentence is unsupported. This is not recorded as a contradiction, because no mutually exclusive value for the same field is present.
- Input an empty array by paper and check the downstream chain. Observe whether the output chain runs through. Do not assert that a fabricated rule will always return an empty array, and do not assert that an empty array stops the chain.