What the documents look like
Multi-store matrix work draws on four document families: store authorization lists, price control tables, channel-conflict complaint forms, and monthly sales reports. Sources are spread across platform merchant backends, ERP channel pricing, brand anti-counterfeit systems, and distributor reports. Update cadence is daily and monthly. For product and marketing copy drafting, the price control table determines the price range a given product may display in a specific store, the store authorization list determines whether a store holds sales eligibility for a brand, and the complaint form records channel disputes that have already occurred. The retrieval step passes these materials to the drafting chain as reference content. Whether a proposed sentence holds depends on whether the material states the subject, object, status, and effective scope. 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.
Constraints at the retrieval step
Retrieval results supply reference content only and do not perform authorization decisions. Semantic recall and the minimum relevance value are similarity filters and are unrelated to store eligibility or price permissions. Hybrid retrieval merges multiple recall paths through RRF; this is a result-merging process, not a separate switch. Question optimization performs coreference resolution and question expansion and adds one model call. Image retrieval requires image input and matching model capability, and reranking requires a rerank model enabled by configuration; these optional paths do not all execute on every request.
Hypothetical drill sample: the source material reads "Authorization list states: Store A Brand Flagship is approved to sell all A-brand products currently on sale during the period January 1, 2025 to December 31, 2025" and "Price control table states: Store A Brand Flagship's listed price range for model A-100 is 199 yuan to 259 yuan." The proposed sentence to be judged is "Store A Brand Flagship sells model A-100 at 259 yuan, the brand-authorized lowest price." The judgment requires the same subject, the same period, and the same field: the material states a listed price range whose upper bound is 259 yuan and does not state the phrase "lowest price," so the proposed sentence is unsupported. The material does not state the supply period for A-100; quote validity and supply period are different facts, and the evidence gap remains.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate the store authorization list and the price control table | Verify both use the same vector model; exclude any whose model identifier differs |
| Search mode | Start with hybrid retrieval | Compare recalled content against semantic retrieval and full-text retrieval using the same test question |
| Reference cap | Set provisionally to 2000 tokens | Test with actual materials of different lengths to see whether reference content is complete, then adjust |
| Minimum relevance | Set provisionally to 0.4 | Label relevant and irrelevant samples first, then compare what remains after filtering |
| Question optimization | Start enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled per configuration | Verify a rerank model is available; compare rerank scores with vector scores |
These values are hypothetical starting points for this trial, used for comparison against adjusted values, and do not constitute defaults. The minimum relevance value under search filtering takes effect only during semantic retrieval or when result reranking is used.
Three common errors
- Treating a retrieval hit as established authorization. Input is the proposed sentence and the authorization list. If retrieval returns an authorization list fragment, compare whether the material states that the store holds brand sales eligibility for the period the draft refers to. If the material states only part of the period, or states the brand without stating the model, the result is unsupported or pending verification; a retrieval hit by itself does not constitute authorization confirmation.
- Treating the relevance threshold as an entity or permission filter. Input is the minimum relevance value and the returned content. Compare whether the returned content includes records for other stores or other brands. The threshold affects similarity filtering only; it does not exclude unrelated business or unauthorized records, and the subject field must be checked separately.
- Associating knowledge bases across different vector models. Input is the vector model identifiers of multiple knowledge bases. Compare whether the identifiers match. When identifiers differ, exclude the non-matching ones and record the actual observation; do not run cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the source material and the proposed sentence from the hypothetical drill sample above, and compare the price wording in the proposed sentence against the listed price range stated in the price control table. When the proposed sentence contains the "lowest price" wording that the material does not state, judge it unsupported and record that result.
- Input a short question containing only "can this store sell it," and compare whether question optimization completes the subject in the retrieval terms before and after. If the subject remains unclear or is altered after optimization, the question is pending verification; record the optimization result.
- Input a paper-level empty result
[], and check whether the downstream drafting chain continues to execute with a reference content length of 0. When reference content is empty, the downstream design should accept that output; do not infer from this that the retrieval rule necessarily returns an empty array.
All three verification items address the retrieval and recall step in product and marketing copy drafting, checking whether the proposed sentence is supported by the material, and do not involve enterprise identity certification or other tasks.