What the documents look like
Price and display auditing happens during day-to-day operations, so the material rolls forward weekly and monthly. The supervision system produces store-visit photographs. The POS price list records the actual selling price in each store. The headquarters promotion calendar states the price and display position that should be executed during a given period. Store cameras supply on-site footage. The working documents are store-visit photographs, price-tag lists, display standards and rectification notices. Together these materials answer one concrete question: on a given date, at a given store, does a given product's price tag and display position match the requirements of the promotion period in force. The order-fulfilment and after-sales question-answering stage needs to retrieve citable passages from these materials in order to answer a verification request raised by a store or a regional supervisor. The update frequency itself does not change the retrieval function; the retrieval action works against the current version already stored in the knowledge base.
The eight-dimension brief places this category inside chain stores and franchising, with the feature being knowledge-base retrieval and recall. The brief marks its own confidence as broadly indicative. Where a detail is not supplied, it stays unknown rather than being filled in from a neighbouring category.
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 built on different vector models cannot be associated together. At operation time, check the vector model identifier first, exclude the ones that do not match, and record what was actually observed. Do not run a cross-model retrieval in order to compare missed recalls.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Compare the three against the test question rather than assuming one mode suits every kind of material. The citation limit is measured in tokens. Test with long and short real materials to see whether content is truncated, then adjust the limit. The minimum relevance score takes effect only under semantic retrieval or when result re-ranking is used; label relevant and irrelevant samples before adjusting the threshold. Question optimisation performs coreference resolution and question expansion. Compare the original question with the optimised question and check whether the subject, store or date has been altered. Semantic recall and the relevance score do not carry out permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to any external system. RRF is a fusion process for mixed results, not an on/off switch.
Images participate in retrieval only when an image input is present and the model capability supports it. Re-ranking runs only when a re-rank model is available and is enabled according to the setting. The optional paths shown in the overall retrieval flow do not all execute on every request.
Hypothetical paper exercise sample. A store-visit photograph carries the text note: "10 March 2025, a store in Zone A, own-brand tissues displayed on the third shelf, price tag shows 9.9 yuan." A headquarters promotion calendar document states: "1 March to 15 March 2025, own-brand tissues execute 8.8 yuan, displayed on the second shelf." A draft sentence awaiting judgement reads: "On 10 March the store's tissue price tag and display both meet the promotion period requirements." Judgement requires the same subject, the same period and the same field. The two materials hold mutually exclusive values on the price field (9.9 yuan against 8.8 yuan) and mutually exclusive values on the display-shelf field (third shelf against second shelf). The draft sentence agrees with neither material and is unsupported; it cannot be judged as compliant. If only the store-visit photograph exists and the promotion calendar document is absent, the price and display requirements are missing and the evidence gap is recorded.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Verification method |
|---|---|---|
| Associated knowledge bases | Store-visit photograph notes, price-tag lists, display standards, promotion calendar | Check each base's vector model identifier and exclude mismatches |
| Search mode | Hybrid retrieval | Run the sample question under semantic, full-text and hybrid, compare recalled passages |
| Citation limit | Hypothetical starting value 2000 tokens | Test once with a long calendar document and once with a short price-tag list, observe truncation |
| Minimum relevance score | Hypothetical starting value 0.4 | Label relevant and irrelevant passages in the sample, observe what remains after filtering |
| Question optimisation | Enabled | Compare original and optimised question, check subject, store and date for alteration |
| Result re-ranking | Enabled per setting | Confirm the re-rank model is available, observe the 0-1 score and ordering change |
The starting values above are chosen for this paper exercise only, so that filtering and truncation behaviour can be compared against a stated reference. They are not documented defaults.
Three common errors
- Treating the minimum relevance score as a permission or entity filter. Input: the sample question "Zone A store, 10 March, tissue price tag." Comparison: whether returned passages contain only that store and that date. Observable result: if passages from other stores or other dates appear, the relevance threshold is not performing isolation, and a separate permission implementation is required.
- Forcing knowledge bases with different vector models into one association. Input: two bases whose vector model identifiers differ. Comparison: whether the node permits selecting both. Observable result: if joint association is not permitted, record that limitation and stop; do not invent missed-recall or redundancy behaviour.
- Treating a rectification notice awaiting review as completed rectification. Input: a rectification notice marked "awaiting review." Comparison: the notice's status field against "completed" in the draft sentence. Observable result: while the status is awaiting review, the draft sentence is unsupported and the evidence gap is recorded.
How to verify the setup
- Verification one. Input the store-visit photograph note and the promotion calendar document from the hypothetical sample above. Compare the price field 9.9 yuan against 8.8 yuan and the shelf field third against second. Observable result: retrieval returns both source passages, and the draft sentence "both meet the promotion period requirements" agrees with neither material, so it is judged unsupported.
- Verification two. Input a query containing only the store-visit photograph note, with no promotion calendar document. Compare whether any promotion-requirement field appears in the retrieved results. Observable result: the promotion requirement is missing, the evidence gap is recorded, and no compliant or non-compliant judgement is issued.
- Verification three. Input an empty citation array [] and check whether the downstream question-answering chain still runs. Observable result: the output chain runs through with a citation length of zero. This does not assert that a fabricated-rule query necessarily returns []. All three verifications are paper checks; no retrieval, approval or rectification action has been executed.