What the documents look like
Marketplace promotion registration and price management centres on price, inventory and campaign rhythm. Promotional wording review sits in a secondary position. The working materials are the campaign registration form, the promotional price approval sheet, the inventory lock sheet and the campaign review sheet, refreshed weekly or per campaign. When a draft sentence states a lowest take-home price, a locked quantity, a registration status or an effective period, that sentence has to line up field by field with those materials. Knowledge base retrieval and recall does one job here: it links a draft sentence to campaign materials already stored in the knowledge base and returns citable reference content. The references carry text evidence only. Whether a price approval passed, whether inventory was actually locked, and whether a registration was accepted by the platform remain separate facts owned by their own business stages.
Constraints at the retrieval step
A knowledge base is made of library, collection and data. The smallest search unit is the library; a search runs across the whole library, and collections only classify data. 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. Reference content is output as an array, and its length can be zero, so the output path still completes when nothing is found.
One retrieval combines question optimisation, semantic retrieval, full-text retrieval or hybrid retrieval, image retrieval, RRF fusion, ReRank, similarity filtering and reference cap trimming. These paths run according to the actual input and settings: image retrieval needs a corresponding image input and model capability, and reranking needs an available model enabled by configuration. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold; it does not isolate permissions, entities or business ownership. The retrieval step does not decide whether a given price record belongs to the current store, and it does not decide whether an approval sheet has taken effect.
Hypothetical paper exercise sample, three source sentences for cross-checking.
Sentence one: "Campaign name: 618 home appliance special; registered product: one model of rice cooker; submitted campaign price: 299 yuan; registration status is submitted, awaiting platform review."
Sentence two: "Promotional price approval sheet: one model of rice cooker, campaign price 299 yuan; approval status: finance reviewed, sales lead signature pending; applicable period: 1 June to 20 June."
Sentence three: "Inventory lock sheet: one model of rice cooker, 500 units locked; lock period 1 June to 20 June; lock status: registered, warehouse confirmation pending return."
Draft sentence to be judged: "This product's campaign price has passed approval, inventory of 500 units is locked for the campaign period, and the registration has been approved by the platform."
That draft sentence contains three assertions beyond what the materials support. The approval sheet is at the signature-pending stage, the lock sheet is at the return-pending stage, and the registration status is awaiting review. All three are unfinished states.
Configuration and implementation checks
| Configuration or check item | Value or action on this page | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases using the same vector model that hold the campaign registration form, promotional price approval sheet, inventory lock sheet and campaign review sheet | The node requires the same vector model; check the vector model identifier first and exclude any that differ |
| Search mode | For this hypothetical trial starting point, select hybrid retrieval | The draft sentence contains both a product name and a price figure, so semantic and keyword clues coexist; compare item by item against semantic-only and full-text-only results |
| Reference cap | For this hypothetical trial starting point, set 1200 tokens, because a single record in these four material types is of medium length and needs to hold price, status and period fields | Compare against 800 and 1600 to see whether reference content truncates key status fields |
| Minimum relevance | For this hypothetical trial starting point, set 0.35 | First label relevant and irrelevant samples by hand, then compare against 0.2 and 0.5 for false recall and missed recall |
| Question optimisation | Enable, and check whether the optimised question keeps product name, price and status as subjects | Compare the original question with the optimised question and confirm the subjects were not rewritten |
| Result reranking | Enable according to whether the current deployment has a rerank model | After enabling, compare how rerank scores and vector scores order the same candidate set |
Three common errors
- Reading the registration status "submitted, awaiting platform review" as registration approved. The input is sentence one; the comparison target is "the registration has been approved by the platform" in the draft sentence; the observable result is that the two status values do not match, and the draft sentence is unsupported by the material.
- Reading the approval sheet status "signature pending" as approval complete. The input is sentence two; the comparison target is "has passed approval" in the draft sentence; the observable result is that the signature step is not finished, so an approval conclusion cannot stand.
- Reading the inventory lock sheet status "registered, warehouse confirmation pending return" as inventory locked. The input is sentence three; the comparison target is "inventory of 500 units is locked" in the draft sentence; the observable result is that the lock status field is still at an intermediate stage, so the locked quantity cannot be written into copy as an accomplished fact.
How to verify the setup
- Input sentence one together with the draft phrase "the registration has been approved by the platform", compare the registration status field value, and when the reference content returns "submitted, awaiting platform review" while the draft still says "approved", judge that sentence unsupported. If the reference content is an empty array, record that the corresponding material was not retrieved and keep the evidence gap open.
- Input sentence two together with the draft phrase "has passed approval", compare the approval status field, and when the reference content returns "signature pending", judge that the approval-complete conclusion does not hold. If a different approval sheet is returned, check whether the product name and applicable period match; if they differ, exclude it.
- Input sentence three together with the draft phrase "inventory of 500 units is locked", compare the lock status field, and when the reference content returns "warehouse confirmation pending return", judge that the lock fact still needs verification. If retrieval does not hit that sheet, record it as unsupported and do not infer whether the warehouse has confirmed.
All three items above are paper cross-checks and were not executed in a real system. Retrieval results provide text evidence only; registration acceptance, approval taking effect and inventory lock confirmation are independent facts belonging to their respective business stages.