What the documents look like
Douyin storefront shelf operations depend on product cards and store experience scores for traffic, and the daily working material sits in four document groups: product qualification pages, experience score details, shipping violation appeals, and store announcements. Copy drafting for products and marketing draws its citable wording from these documents. Any promotional sentence, time-limit sentence, or scope-of-application sentence in a draft needs a corresponding subject and status in the retrieved source text. The material updates daily, and order-related records are produced per order, so retrieval works against text that keeps changing; the same field can hold different values at different times.
Because the copy task is about wording support, the retrieval step only needs to answer whether a proposed sentence can be traced to source wording on the same subject, the same period, and the same field. It does not decide whether a store is authorized, whether a product may be published, or whether a qualification is valid.
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 with different vector models cannot be associated together. The vector model identifier is checked first, and non-matching bases are excluded; no cross-model retrieval is run to compare missed recalls.
The node outputs cited content as an array, and the length can be 0. When nothing is retrieved, that output path still completes. Semantic retrieval, full-text retrieval, and hybrid retrieval are chosen by material type: identifiers, model numbers, and exact queries fit full-text retrieval better, while natural-language questions fit semantic retrieval better. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. The citation limit is measured in tokens. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Retrieval results only indicate content relevance; they do not carry permission or entity isolation, and they do not judge whether copy may be published externally.
Hypothetical drill sample: a Douyin store merchant on 10 June 2025 organizes store announcements and product qualification pages to draft a product card sentence. The source text for comparison is: "This store has provided 7-day no-reason returns for signed orders since 1 June 2025; return requests must be submitted within 7 days of signing." A second sentence reads: "The quality inspection report number for a summer short-sleeve item is A-2025-0610, the report issue date is 20 May 2025, and the applicable batch is the May 2025 production batch." The draft sentence to be judged is: "All products in this store support 30-day no-reason returns and have passed national authoritative quality inspection." The subject scope, the time limit, and the inspection conclusion in this sentence all need item-by-item comparison with the source text above.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model base containing store announcements and product qualification pages | Compare vector model identifiers against other bases and exclude mismatches |
| Search mode | Run semantic retrieval and full-text retrieval separately on the same question | Compare cited content returned by each mode and observe the difference |
| Citation limit | Start at 2000 tokens | Compare against actual material of different lengths and observe whether citations are truncated |
| Minimum relevance | Start at 0.4 | Compare against labeled relevant and irrelevant samples and observe the filtering effect |
| Question optimization | Enable for follow-up questions containing references; leave standalone short sentences off | Compare against retrieval on the original question and check whether the subject was altered |
| Result re-ranking | Enable when a re-ranking model is available and candidates are numerous | Compare candidate order with and without re-ranking and observe score changes |
These values are hypothetical starting points for this drill and are not official defaults.
Three common errors
- Reading "7-day no-reason returns for signed orders" in the source text as all products and 30 days. The draft sentence exceeds the source in both subject scope and time limit, so it is unsupported.
- Treating quality inspection report number A-2025-0610 as equivalent to "national authoritative quality inspection." The source states only the report number, issue date, and applicable batch; it does not state that conclusion, so this is an evidence gap.
- Treating minimum relevance as a permission or entity filter, assuming that raising the value excludes unrelated stores or unauthorized records. Retrieval itself does not carry that responsibility.
How to verify the setup
- Input the source text and the draft sentence from the hypothetical sample, and compare whether the subject scope, time limit, and inspection conclusion in the draft sentence can be found in the source with the same subject, the same period, and the same field. The draft sentence is supported only when the source explicitly states the same value; when the source does not state it or provides only related-stage evidence, record it as unsupported or pending verification.
- Input a follow-up question containing a reference, retrieve once with question optimization enabled and once disabled, and compare whether the cited content returned in the two runs points to the same subject. When the optimized run alters the subject, record that optimization result as unusable for this drafting task.
- Input a paper empty-result scenario and check how the downstream design handles a cited-content array of length 0, observing whether the path continues. Do not claim that a fictional rule retrieval necessarily returns an empty array; record only the output actually observed.