What the documents look like
Marketplace advertising adjusts bids against live traffic and conversion signals, so the working material changes on a daily or hourly rhythm rather than waiting on a platform campaign review. The retrieval corpus for this category is built from campaign plans, keyword reports, audience package descriptions, and conversion attribution tables. Keyword reports and attribution tables carry exact fields such as identifiers and model codes, and cross-border material adds parallel wording in more than one language. A single retrieval mode struggles to cover natural-language questions and exact-code questions at the same time. Planners, brand marketing teams, and agency media buyers use recalled content to reconcile attribution and to trace conversion breakpoints, so incomplete recall directly affects bid adjustment and budget reconciliation.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the vector model identifier is checked first, non-matching bases are excluded, and the actual observation is recorded. Retrieval is not authorization: recalled content supports reconciliation and does not replace business approval, database writes, or external synchronization.
Semantic retrieval works from vector distance and suits natural-language and cross-language questions. Full-text retrieval works from keyword hits and suits identifiers, model codes, and proper nouns. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a separate switch. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold, not a sorting rule, and it does not perform permission or entity isolation. The reference cap is measured in tokens and limits how much content a single search may cite. Question optimization performs coreference removal and question expansion and adds one model call. Image retrieval requires image input and matching model capability, and reranking requires an available model enabled under the configuration; optional paths drawn in the documented flow do not mean every request executes all of them. Empty reference output is an array whose length may be 0, and the downstream link still runs.
Configuration and implementation checks
Hypothetical drill sample: the source material reads, "From 1 June 2024 to 30 June 2024, plan QH-2406 ran on the Qianchuan channel with a daily budget of 800 yuan and a target ROI of 2.5; on 15 June the plan was paused because the creative failed review." The draft sentence to be judged reads, "Plan QH-2406 ran for the whole of June and met its ROI target." This sample checks whether retrieval can recall plan status and period information.
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the campaign material base that uses the same vector model | Check each base's vector model identifier, exclude non-matching ones, then associate |
| Search mode | Test natural-language questions with semantic retrieval first, then test code questions with full-text retrieval | Compare recalled content for the two question types; do not assume one mode fits all material |
| Reference cap | Set 2000 tokens for this drill | Test with short and long material, compare against the tokens a complete passage needs, then adjust |
| Minimum relevance | Set 0.4 for this drill | Label relevant and irrelevant samples first, then move the threshold; it applies only in semantic retrieval or reranking |
| Question optimization | Leave off for now | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Leave disabled for now | Confirm the rerank model is available and enabled under the configuration, then compare ordering changes |
Three common errors
- Associating knowledge bases that use different vector models. Input: two bases with different vector model identifiers. Comparison: whether the identifiers match. Observable result: when they do not match, the bases are not associated, and only that observation is recorded.
- Treating minimum relevance as a permission or entity filter. Input: one record outside the current business scope. Comparison: whether that record is excluded by the relevance threshold. Observable result: the threshold filters by similarity only and does not isolate permissions.
- Describing RRF as a separate switch that can be turned on. Input: a hybrid retrieval configuration. Comparison: whether such a switch exists in the configuration. Observable result: the documentation describes RRF only as a fusion process and provides no separate switch.
How to verify the setup
- Using the hypothetical source material and draft sentence, input the retrieval question "What was the delivery status of plan QH-2406 in June?" Compare whether the recalled content includes the pause date and the pause reason. The condition is met when recalled content matches the source material; if it is not recalled, record it as pending verification and do not treat it as a contradiction.
- Input the code question "QH-2406" and compare the recall difference between full-text retrieval and semantic retrieval. The condition is met when full-text retrieval hits the code; a semantic miss is not treated as a failure.
- Input the explicit paper input
[]for an empty result and compare whether the downstream link continues to execute. The condition is met when the link runs through; no claim is made that a fictional rule search necessarily returns[].