What the documents look like
Platform advertising adjusts bids against traffic and conversion in near real time, and it does not depend on platform campaign review the way promotional registration does. The working materials are delivery plans, keyword reports, audience package descriptions and conversion attribution tables. These update at daily and hourly cadence. The people involved are delivery specialists, brand marketing teams and agency media buyers, and account or material handover between them happens regularly.
For this page the relevant point is narrower. A retrieval result reflects only content that has been stored in the knowledge base and has completed training. Retrieval itself does not decide who may view or change a given delivery record. The knowledge base search node outputs cited content as an array, and the array length can be zero, so the output path still runs when nothing is found.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. Result reranking requires a ReRank model and must be enabled in the settings. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it acts as a filter threshold. It does not perform permission or entity isolation. The citation cap is measured in tokens and limits how much content a single search may cite. Question optimization handles coreference resolution and question expansion and adds an extra model call.
Hypothetical drill sample: one handover material states, "From 1 March 2025, authority to adjust the daily budget of brand A's Qianchuan account transfers from buyer Jia to buyer Yi, with the handover period ending 31 March." A second material states, "On 15 March, buyer Yi requested an increase of the daily budget from 800 yuan to 1,200 yuan; the request status is pending review." The draft sentence to be judged is, "Buyer Yi completed the daily budget adjustment on 15 March." Checking must cover both materials, comparing subject, object, status and effective scope. A pending request does not amount to execution, and the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the delivery handover repository that uses the same vector model | First check the vector model identifier of each repository, exclude those with a different identifier, and do not run a cross-model retrieval comparison |
| Search mode | Hybrid retrieval | Run the same test question through semantic, full-text and hybrid modes and compare the recalled content; do not assume one mode suits all materials |
| Citation cap | Measured in tokens, starting at a length that covers one complete handover statement | Test with actual materials of different lengths to see whether content stays complete, then adjust the cap |
| Minimum relevance | Starting value 0.5 | Label relevant and irrelevant samples first, compare what remains after filtering, then adjust the threshold |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether subject or object was altered incorrectly |
| Result reranking | Enabled per settings, with the ReRank model confirmed present | Compare candidate order before and after reranking and check how the score works with minimum relevance |
Three common errors
- Treating a retrieval hit as confirmation of authority. Input "buyer Yi daily budget adjustment," compare the retrieved content with the authorization status in the handover materials, and observe whether a pending request is written up as a completed adjustment.
- Treating minimum relevance as a permission or entity filter. Input the same question and adjust the threshold, compare whether returned content still includes records of other brands or other subjects, and observe whether the threshold is mistakenly believed to exclude unrelated business.
- Treating RRF as a switch that can be turned on. Input a hybrid retrieval request, compare the multiple result paths before and after merging, and observe whether RRF is described as an independent configuration item.
How to verify the setup
- Input the two material statements and the draft sentence from the hypothetical sample, and compare subject, object, status and effective scope. If the draft sentence says the adjustment was completed while the materials only record a pending review, the condition is not met.
- Input the same test question and compare the recalled content of semantic, full-text and hybrid retrieval. If the hybrid result contains hits from multiple paths merged through RRF, the condition is met. If retrieval returns nothing, record the gap and do not infer a cause.
- Input an empty-citation scenario and check whether the downstream design accepts an array output of length zero. If downstream execution continues, the condition is met. Do not claim that a fictional rule retrieval necessarily returns an empty array.