What the documents look like
Marketplace advertising and on-site traffic work draws on four recurring document families: campaign plans, keyword reports, audience package descriptions, and conversion attribution tables. The public brief describes update frequency as daily or hourly, and the typical practitioners as advertising specialists, brand marketing teams, and agency media buyers. For store service and internal training, retrieval requests usually arrive as immediate questions from frontline staff about campaign rules, bid conventions, or attribution conventions.
A practical consequence of this document mix is that the subject, target object, and effective scope of one item are often spread across separate files. A campaign plan may state a bid and a date range for one account, while a handover request in another file may propose a different bid and a later period. The retrieval step has to gather content that shares the same subject, the same period, and the same field so that a human can compare them. Retrieved content is candidate reference material only. It does not constitute authorization, approval, or execution.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Semantic retrieval, full-text retrieval, and hybrid retrieval suit different document shapes. Queries built from numbers, model codes, and proper nouns are better served by full-text or hybrid retrieval, while natural-language questions are better served by semantic retrieval or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and Rerank performs a second ordering of candidate text. Rerank requires a model and is enabled according to settings. Image retrieval requires image input and corresponding model capability, so it is an optional path rather than a step that runs on every request.
The reference cap is measured in tokens. The minimum relevance score is a filter threshold that takes effect only in semantic retrieval or when result re-ranking is used. It is not a final ordering rule and does not exclude unrelated businesses or unauthorized records. Question optimization performs coreference resolution and question expansion before retrieval and adds an extra model call. Retrieval does not perform authorization or entity isolation.
A hypothetical paper exercise sample is given here for later comparison. The training material text reads: "In this store's March campaign plan, the keyword 'spring outerwear' bid is adjusted to 1.20 yuan, the effective period is March 1 to March 31, and the applicable scope is limited to this store's account." A separate handover request reads: "Please change that keyword bid to 1.50 yuan and continue using it in April." The judgment question is whether the request is consistent with the source text on bid, period, and applicable scope. The source text does not state any April arrangement and does not name an approver.
Configuration and implementation checks
| Configuration item | Relevant setting or check for this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases that use the same vector model | The sample material must be retrievable within a single vector-model knowledge base |
| Search mode | Compare semantic, full-text, and hybrid retrieval against the test question | Use "spring outerwear bid" to test a mixed wording |
| Reference cap | Unit is tokens; test completeness with materials of different lengths | Check whether the March plan and the handover request return in the same segment |
| Minimum relevance | Takes effect only in semantic retrieval or result re-ranking | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result re-ranking | Requires a model and is enabled according to settings | Observe ordering changes when candidates are numerous |
Three common errors
- Treating a retrieval hit as authorization. In the sample, the handover request mentions continuing in April, while the source text states only March 1 to March 31. Retrieving that request does not mean an April change has been approved. The check compares the period value in the same field and records the unstated part as pending verification.
- Associating knowledge bases that use different vector models. The documents require associated knowledge bases to share the same vector model, so knowledge bases with different models cannot be associated together. The check inspects the knowledge base identifier first, excludes non-conforming bases, and records what is actually observed without inferring runtime behavior for cross-model retrieval.
- Treating the minimum relevance score as a permission or entity filter. That value is a filter threshold that takes effect only in semantic retrieval or result re-ranking, and it does not exclude unrelated businesses or unauthorized records. The check compares referenced content before and after a threshold adjustment, and handles permission questions separately.
How to verify the setup
- Input the source text "In this store's March campaign plan, the keyword 'spring outerwear' bid is adjusted to 1.20 yuan, the effective period is March 1 to March 31, and the applicable scope is limited to this store's account" together with the draft sentence "Please change that keyword bid to 1.50 yuan and continue using it in April." Compare the bid, period, and applicable scope in the same fields. The source text gives a bid of 1.20 yuan, a period of March 1 to March 31, and a scope limited to this store's account; the draft sentence gives a bid of 1.50 yuan and points to April. The bid and period differ from the source text, and the April arrangement is unsupported. Record it as pending verification rather than as approved.
- Input the query "spring outerwear bid 1.20" and observe the returned reference content under semantic retrieval, full-text retrieval, and hybrid retrieval separately. Compare whether the returned content contains the paragraph where the source text sits. If the returned content is an array with length 0, the query returned no content; record it as an empty result and do not claim that the rule always returns an empty array. If the returned content contains the source paragraph, check whether subject, period, and applicable scope are complete.
- Input the handover request "Please change that keyword bid to 1.50 yuan and continue using it in April" and compare the request text with the source text inside the retrieval results. Observe whether retrieval returns both materials. If both are returned, check whether the same field carries mutually exclusive values. The source period is March and the request points to April, which are different periods and do not form a same-field contradiction. The request is pending review, which does not indicate that operating permission has been granted and does not indicate that the change has been executed.