What the documents look like
Brand member private-domain operations draw on CRM systems, enterprise messaging, POS purchase records and mini-program behavior data. The working documents are member tag tables, outreach plans, coupon rules and privacy authorization records. Copy drafting for products and marketing takes member tier, benefit scope, coupon usage conditions and authorization status from these materials. Member tag tables and outreach plans typically refresh daily or weekly; the retrieval function itself does not depend on that rhythm, only on the index state after data has been stored.
The category difference that matters here is that copy for this group carries benefit promises and authorization boundaries. A retrieval result has to show the drafter the source wording, so the drafter can judge whether a proposed sentence is supported. Knowledge base search treats the library as the smallest search unit; a collection is only a classification layer and does not affect search results. One retrieval run can combine question optimization, semantic retrieval or full-text retrieval, RRF fusion, ReRank, similarity filtering and a reference token cap, and finally outputs a reference content array whose length can be zero.
Constraints at the retrieval step
Search parameters set the candidate content range. They do not set business permissions. The minimum relevance value is a filter threshold and takes effect only under semantic retrieval or when result reranking is used. It does not act as a filter for member tier or authorization status. The reference cap is measured in tokens and controls how much content a single search may cite. Coupon rules and authorization records vary in length, so content completeness needs to be tested against both short and long material.
Question optimization performs coreference resolution and question expansion, calls the model an extra time, and can change the subject of the question. Semantic retrieval suits natural-language phrasing; full-text retrieval suits numbers, model names and proper nouns; hybrid retrieval merges the two paths through RRF; ReRank requires a model and runs according to configuration. Image retrieval requires an image input and the corresponding model capability, and is not a path that every request executes. When several knowledge bases are associated, they must use the same vector model. Libraries with different vector models cannot be associated together; the model identifier is checked first and non-matching libraries are excluded.
Hypothetical drill sample
Source material for comparison (fictional, used only to demonstrate the comparison method on this page):
"The March 2025 member benefit table states: Silver members may claim one coupon of 30 yuan off a 200 yuan purchase during their birthday month. The coupon is valid for 30 days from the claim date. Applicable stores are directly operated stores in the East China region."
Draft sentence to be assessed (fictional):
"Silver members may claim a 30 yuan off 200 yuan coupon during their birthday month, and the coupon is valid at stores nationwide."
The object of comparison is the applicable store scope of the coupon. In the source material the same field reads "directly operated stores in the East China region"; the draft sentence writes "stores nationwide". These are mutually exclusive values for the same field, which constitutes a contradiction. The source material does not state a limit on the number of claims, and the draft sentence does not address it, so that point is unsupported. The source material does not state any authorization record, so privacy authorization status is an evidence gap and requires separate checking.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Member benefit rules library and coupon rules library, two libraries | Both libraries must use the same vector model; check the model identifier first and exclude non-matching libraries |
| Search mode | Hybrid retrieval | Coupon rules contain exact wording such as "30 yuan off 200 yuan" alongside natural-language conditions; compare hit content against semantic retrieval alone and full-text retrieval alone |
| Reference cap | Start with a token value that covers the full text of one coupon rule | Test with coupon rules and authorization records of different lengths and compare whether reference content is truncated |
| Minimum relevance | Start with a middle value | Label relevant and irrelevant samples first, then compare whether the coupon applicable-store field survives filtering |
| Question optimization | Start with it off | The subject of this sample question is already clear; compare whether the question subject is rewritten when it is on versus off |
| Result reranking | Enabled according to configuration | A rerank model is required; compare how the rerank score and the vector score affect candidate order |
All values above are starting points for this hypothetical trial and do not constitute official defaults.
Three common errors
- Treating minimum relevance as a filter for member tier or authorization status. The input is a coupon-rule query for a silver member; the comparison object is the reference content before and after filtering; the observable result is that the relevance threshold affects similarity scores and does not change how the member tier field is selected.
- Associating knowledge bases that use different vector models and then inferring runtime behavior. The input is two libraries with different model identifiers; the comparison object is the model identifier field; the observable result is that libraries with inconsistent identifiers cannot be associated together, and cross-model retrieval behavior is unsupported.
- Treating RRF as a switch that can be turned on. The input is a hybrid retrieval configuration; the comparison object is where RRF is described in the documentation; the observable result is that RRF is a fusion process and no corresponding switch exists among the configuration items.
How to verify the setup
- Input the hypothetical source material and draft sentence above; compare the coupon applicable-store field "directly operated stores in the East China region" with the draft sentence's "stores nationwide"; the observable result is that these are mutually exclusive values for the same field, judged as a contradiction, and the draft sentence needs revision.
- Input the hypothetical source material and draft sentence above; compare the coupon claim-count limit field; the source material does not state this field; the observable result is that this point is unsupported and requires additional material before a judgment can be made.
- Input an explicit paper empty result []; compare how the downstream design handles an empty reference array; the observable result is that the output path still completes when reference content length is zero. This does not claim that a search of the fictional rules must return [].