What the documents look like
Campaign plans, keyword reports, audience package descriptions and conversion attribution tables form the main document surface for this category. Advertising adjusts bids against traffic and conversion in real time, so these materials are updated on daily and hourly cycles. A retrieval step reads whatever was already ingested and trained at a given moment; it does not reach into a live bidding system.
The same product may appear under different names in a plan, a report and an attribution table. Whether the query wording and the indexed content line up directly affects recall quality. The people handling these materials are campaign specialists, brand marketing teams and agency media buyers. Retrieval output is quoted into product selection question answering; it does not decide authorization.
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 on different vector models cannot be associated together; check the model identifier first and exclude those that do not match.
Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval merges the two paths through RRF, and RRF is a fusion process rather than a switch that can be turned on. The minimum relevance setting takes effect only under semantic retrieval or when result re-ranking is used. It is a filter threshold and does not isolate records by permission, entity or business ownership.
The citation limit is measured in tokens and controls the length of returned content. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval. Image retrieval requires corresponding image input and model capability. Re-ranking requires a model and runs according to its setting. Neither runs automatically on every request. Quoted content is returned as an array, and its length can be zero.
Configuration and implementation checks
Hypothetical drill sample. Source text for comparison: "Qianchuan plan A-17 ran from March 1 to March 7, with a daily budget of 800 yuan, and its status is paused." "Keyword report KW-03 covers March 1 to March 7, with 1,200 clicks and 36 conversions." Draft sentence to be assessed: "A-17 ran throughout the whole of March with no budget adjustment." Handover request: "Please confirm whether A-17 can continue to scale up." The source text covers only March 1 to March 7 and records a paused status, so the draft sentence's "throughout the whole of March" is not supported, and the handover request concerns scaling authorization. Retrieval and forms do not constitute authorization execution.
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the campaign materials base | Check its vector model identifier against the node requirement and exclude bases with a different model |
| Search mode | Run semantic retrieval and full-text retrieval in parallel | Compare against semantic retrieval alone and observe whether identifier-type questions improve |
| Citation limit | Set to a token amount covering one complete campaign description | Compare against a shorter starting point and observe whether campaign start and end dates are truncated |
| Minimum relevance | Start at a middle value | Compare against lower and higher values and observe whether unrelated fragments are brought in |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare against the disabled state and check whether the subject is altered incorrectly |
| Result re-ranking | Enable when candidates are numerous | Compare against unranked results and observe whether ordering moves closer to the question |
Three common errors
- Treating a retrieval hit as authorization. Input "Please confirm whether A-17 can continue to scale up." The comparison object is the retrieval output against the authorization record. A returned campaign description does not constitute approval to scale up, and approval requires a separate downstream system implementation.
- Treating partial-period evidence as a whole-period conclusion. Input the draft sentence "A-17 ran throughout the whole of March." The comparison object is the source text range of March 1 to March 7. The portion beyond that range is unsupported, and the evidence gap must be recorded.
- Treating a similarity setting as an entity filter. Input a query about an unrelated business. The comparison object is the returned content against the query subject. Minimum relevance filters only low-relevance results; it does not isolate records by store or permission.
How to verify the setup
- Input the two source sentences and the draft sentence "A-17 ran throughout the whole of March with no budget adjustment." Compare the draft sentence against the source text on the period and status fields. The source text records only March 1 to March 7 and a paused status, so the draft sentence exceeds that range and is judged unsupported, with the gap retained.
- Input the handover request "Please confirm whether A-17 can continue to scale up." Compare the retrieval output against the authorization state. The retrieval result does not change the authorization state, scaling requires a downstream approval implementation, and the request is recorded as pending review.
- Input an intent to associate a knowledge base that points to a different vector model. Compare the vector model identifier of that base. When the identifier does not match, exclude the base and record the actual observation, without running a cross-model retrieval to compare missed recall. If retrieval returns an empty array, use the explicit paper input
[]to check whether the downstream chain proceeds, and do not claim that a made-up rule necessarily returns an empty result.