What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so product information and selection questions draw on documents the brand itself produces. Daily sales reports, shift schedules, display photos and inventory transfer orders form the main material, sourced from the POS system, member CRM, ERP inventory and store inspections. These materials update daily or per order, and retrieval faces multiple records for one product code across different dates, stores and fields.
The difficulty in product selection questions is the mixed material format. Text reports carry codes and quantities, display photos carry visual information, and transfer orders carry status fields. Semantic retrieval relies on vector similarity, full-text retrieval relies on keyword hits, and hybrid retrieval merges both paths through RRF. The retrieval step must confirm that associated knowledge bases use the same vector model; image retrieval requires a vision model or a vector model that supports images; reranking requires a ReRank model enabled according to settings. These conditions determine whether retrieval can cover the codes, images and status fields in the material.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together; the operator first checks the identifier and excludes non-matching bases, and does not run cross-model retrieval to compare missed recalls. Reference content is output as an array and its length can be 0, meaning the output path still runs even when no content is found. The downstream design for empty results can be checked with the explicit paper input [], without claiming that a fabricated rule makes retrieval return [].
The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. This value is a filter threshold, not a final ranking rule. After question optimization, hybrid retrieval, image retrieval or result reranking is enabled, final results may pass through multi-path recall fusion and are not strictly ordered by a single vector similarity. The reference limit uses tokens as its unit and controls how much content each search references. Question optimization performs coreference resolution and question expansion, adds one model call, helps retrieval in continuous conversations, and increases overall latency.
Image retrieval requires a corresponding image input and model capability, and reranking requires a model enabled according to settings. The optional paths listed in the overall document flow do not mean every request executes all of them. Retrieval does not handle permission or entity isolation; semantic recall and relevance do not perform authorization decisions. The same-vector-model requirement for retrieval bases does not extend to business registration sites or other external systems.
Configuration and implementation checks
The following configuration concerns a hypothetical drill sample and lists only officially described settings or implementation check items related to this function and primary task.
| Configuration item | Hypothetical starting point for this drill | Reason for selection and comparison method |
|---|---|---|
| Associated knowledge base | Associate one store product information base that uses the same vector model | Check the vector model identifier first, compare it with another base identifier, and do not associate together when they differ |
| Search mode | Hybrid retrieval | The sample contains codes and natural-language questions; compare semantic retrieval and full-text retrieval hits separately |
| Reference limit | Set to a value covering a complete passage of the sample material | Hypothetical starting point for this drill; test with materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Takes effect only in semantic retrieval or result reranking; compare candidate changes before and after filtering |
| Question optimization | Enable for follow-up questions containing references | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | A ReRank model is available and enabled according to settings | Compare ordering before and after reranking and confirm whether candidates are sufficient |
Three common errors
- Treating retrieval results as authorization evidence. In the hypothetical drill sample, a store submits an inventory transfer order request and retrieval hits the draft record of that order; draft status does not indicate approval. The input is the transfer order draft, the compared field is the approval status, and the observed result is that the status field is draft or blank, so the authorization boundary is not met and a separate approval system implementation is required.
- Associating knowledge bases with different vector models together. In the hypothetical drill sample, two knowledge bases use different vector models; the operator checks identifiers first, excludes non-matching ones, and does not run cross-model retrieval to compare missed recalls. The input is the vector model identifiers of the two bases, the comparison is whether they match, and the observed result is that they cannot be associated together when they differ.
- Treating minimum relevance as a ranking rule. In the hypothetical drill sample, minimum relevance is adjusted after hybrid retrieval is enabled; this value only filters low-relevance results and does not determine final ordering. The input is the filter threshold, the comparison is the candidate list before and after filtering, and the observed result is that list order may change through RRF fusion and is not arranged by a single vector similarity.
How to verify the setup
Hypothetical drill sample: a store manager submits a draft inventory transfer order. The material text reads "Transfer order number DB-2024-0317, source store A, destination store B, product code SKU-8821, quantity 12 units, status: draft, submission date 17 March 2024." The sentence to be assessed is "This transfer order has been approved and the transfer can be executed." The material contains the subject, object, status and date, and the status field is draft.
- Input the proposed sentence and the material text, compare "has been approved" in the proposed sentence with the status field "draft" in the material, and observe that when the two do not match, the proposed sentence is not supported and the approval record must be verified.
- Input the retrieval question "transfer status of SKU-8821", compare the reference content returned by retrieval with the material text, and observe that when the reference content contains the status field without rewriting it to another status, the retrieval hit is valid.
- Input an empty-result scenario, use the explicit paper input
[]to check the downstream design, compare whether the downstream handles it as an empty array, and observe that when the downstream path runs without error, the empty-result handling matches the documented description.
These checks describe what to compare and what remains unknown. They do not claim that a test, deployment, approval, refund or other business action actually happened.