What the documents look like
Brand member private-domain operations handle customer data and repeat-purchase outreach, and do not deal with store shift scheduling. Copy drafting for this category draws on four recurring document types: member tag tables, outreach plans, coupon rules, and privacy authorization records. These materials originate in the CRM system, enterprise WeChat, POS purchase records, and mini-program behavioral data, and they update daily or weekly.
For citation tracing, the practical consequence is that a promotional sentence about member tier, coupon threshold, or outreach window must be checked back against the specific material that states the subject, the object, and the validity scope. When two versions of the same material disagree, the citation chain breaks. The update frequency itself does not change the tracing method; what matters is which version was in force for the period the copy describes.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. That array is passed to the answering stage, where each promotional sentence in the draft is compared line by line against the source text. A retrieval score reflects similarity only; it does not perform permission filtering or entity isolation. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch that can be turned on. Reranking requires a model and is enabled according to settings, and image retrieval requires image input together with model capability. These optional paths are listed in the overall flow, and their presence there does not mean every request executes all of them.
Hypothetical drill sample: the source material reads, "From March 1 to March 31, 2025, silver-tier members who spend 300 yuan at designated stores may claim one 50-yuan category coupon; the coupon is valid only in the current month and expires thereafter." A draft sentence reads, "Silver-tier members who spend 300 yuan in March can claim a 50-yuan coupon, valid all year." To judge this, compare whether the subject is the same silver tier, whether the period is the same March window, and whether the coupon validity field matches the source wording. The source states current-month validity only; the draft's "valid all year" is not supported by that field.
Configuration and implementation checks
| Configuration or check item | Use on this page | Value or check method |
|---|---|---|
| Associated knowledge bases | Fix the retrieval scope | Verify selected knowledge bases and material versions; associate only knowledge bases sharing the same vector model |
| Search mode | Match draft sentences to source text | For precise fields such as numbers and coupon names, choose full-text or hybrid retrieval |
| Citation limit | Control total cited tokens | Hypothetical starting point of 800 tokens for this drill, chosen because the sample passage is short; compare with 400 tokens to see whether coverage is complete |
| Minimum relevance | Filter low-relevance results | Hypothetical starting point of 0.5 for this drill; compare with 0.3 to see whether unrelated coupon rules enter the results |
| Citation content output | Carry retrieval results forward | Verify the output array is passed to the answering stage; an empty array is permitted |
| Material validity check | Confirm the version in force | Test sources with current and historical versions; before adoption, verify the validity period |
Three common errors
- Treating the retrieval score as a permission filter. The error appears when a low similarity score is read as excluding unauthorized records. Check by submitting the same question and comparing the returned citations against the authorization scope list, observing whether any record outside that scope appears.
- Treating RRF as a switch setting. The error appears when a configuration screen is searched for an RRF enable option. Check by reviewing the search mode description and observing whether hybrid retrieval results arise from multi-path fusion.
- Judging an unstated field as a contradiction. The error appears when a draft sentence says "valid all year" while the source says only "valid in the current month," and this is recorded as a negation. Check by comparing the explicit value of the same field; where the source does not state an extension of validity, record the item as pending verification.
How to verify the setup
- Enter the hypothetical sample source text and the draft sentence, then compare three fields: subject, period, and coupon validity. The condition is satisfied only when all three fields match the source. Where the draft says "valid all year" and the source states current-month validity only, record the draft as unsupported.
- Enter an explicit paper input of an empty array
[]and check whether the downstream answering stage continues on an empty citation. Do not claim that a fabricated rule retrieval returns an empty array. - Verify the vector model identifiers of the two knowledge bases, exclude any whose identifiers do not match, and then retrieve. Record the actual returned citation entries and their correspondence to the source text. Do not run cross-model retrieval to compare missed recalls.