What the documents look like
Copy drafting for a TikTok Shop US store draws on four document families: product listings, return and refund records, Return Grading evidence, and settlement statements. US local return windows and return grading impose after-sales constraints that differ from shelf-based operations in other markets. Sentences about return promises, time limits, and responsibility allocation need to align with after-sales records and platform rule materials.
The retrieval goal for copy drafting is to let a proposed sentence find support in indexed materials that share the same subject, the same period, and the same field. The search node outputs an array of cited content, and its length can be zero, so the output path still completes when nothing is found. Retrieval output does not perform authorization judgment and does not perform subject isolation.
Materials update daily and per order. Update frequency does not change retrieval configuration; it only affects which material period a person selects during manual cross-checking.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together. Check the vector model identifier of each base, exclude those that do not match, and record what is observed.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Semantic retrieval computes similarity from vector distance and suits natural-language questions. Full-text retrieval suits numbers, model designations, and proper nouns. Hybrid retrieval uses both recall paths and merges them through RRF; RRF is a fusion process, not a switch that can be turned on. The citation limit is measured in tokens and caps how much content one search may cite. Minimum relevance is a 0–1 filter threshold that takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold, not a ranking rule. Question optimization performs coreference resolution and question expansion and makes an additional model call.
Hypothetical paper exercise sample. A merchant knowledge base contains a return policy passage: "Orders on this store's US site may apply for return within 30 calendar days from the day after signed receipt; apparel items must keep tags intact; applications beyond 30 days are handled by the platform according to the return grading result." A settlement statement record reads: "March 2025 settlement cycle, order number US-0001, deduction item is return handling fee, amount as listed on the platform statement." The sentence to be judged is: "All products in this store support 45-day no-reason return after signed receipt." The check task is whether that sentence is supported by the materials above. The materials state no 45-day period and no "all products" scope, so the judgment is not supported.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Hypothetical trial starting value and comparison method |
|---|---|---|
| Associated knowledge bases | Select the bases holding product listings, return and refund records, Return Grading evidence, and settlement statements | First check whether the vector model identifiers of the bases match; exclude those that do not |
| Search mode | Choose semantic, full-text, or hybrid retrieval according to the type of proposed sentence | Use full-text retrieval for sentences containing numbers, semantic retrieval for natural-language sentences, and compare recalled content |
| Citation limit | Control cited content volume in tokens | Set a hypothetical starting value long enough to hold one return policy passage, then compare completeness against longer and shorter materials |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking | Label relevant and irrelevant samples first, then adjust the threshold and observe what remains after filtering |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether subject and object were altered incorrectly |
| Result reranking | Requires a rerank model and must be enabled by configuration | Enable it for proposed sentences with many candidates, then compare rerank scores with vector scores |
Three common errors
- Treating minimum relevance as a permission or subject filter. Minimum relevance is a similarity filter threshold and cannot exclude unrelated businesses or unauthorized records. Check: enter a proposed sentence that is semantically close to the target material but concerns a different subject, observe the returned content, confirm whether that subject's material still appears, and record what is observed.
- Forcing knowledge bases with different vector models into one association. The documentation states that associated knowledge bases must use the same vector model. Check: verify the vector model identifier of each base, exclude those whose identifiers differ, and do not run cross-model retrieval to compare missed recalls.
- Treating RRF as a switch that can be turned on. RRF is the fusion process for hybrid retrieval results and is not a configuration item. Check: inspect the hybrid retrieval configuration interface, confirm whether a switch named RRF exists, and record what is observed.
How to verify the setup
- Check one. For the proposed sentence "All products in this store support 45-day no-reason return after signed receipt" in the hypothetical sample, enter the search terms "return 45 days all products" and compare the returned cited content with the return policy passage. If the returned content contains only the 30-day period and does not contain 45 days, retrieval did not find material supporting that sentence, and the judgment is not supported. The expected result is not claimed to have been tested.
- Check two. For the settlement statement record in the hypothetical sample, enter the search terms "US-0001 return handling fee" and compare the returned cited content with the statement passage. If the returned content contains order number US-0001 and the return handling fee field, the record can be retrieved. If the return is an empty array, check whether the material has finished training; do not claim that a fabricated rule necessarily returns an empty array.
- Check three. For question optimization, enter the short question "What about returns then" in a continuous conversation and compare the optimized question with the subject of the previous turn. If the optimized question completes the subject and object, coreference resolution took effect; if the subject was altered incorrectly, record that observation. The expected result is not claimed to have been tested.
All three checks use the hypothetical paper exercise sample as input and do not claim that any business action was executed.