What the documents look like
Platform advertising adjusts bids against traffic and conversion in real time, so the primary records are campaign plans, keyword reports, audience package descriptions and conversion attribution tables. The data originates from advertising platforms and store transaction records, and the typical operators are media buyers, brand marketing teams and agency traders. The operational consequence of weak source tracing is wasted spend, distorted return-on-investment figures and budget cuts during promotional periods. Operating reports and anomaly reviews have to tie each figure and conclusion back to a specific source document, a specific version and a specific field. Campaign data changes frequently, and hourly updates are common, but version and field-definition checks do not complete themselves because the underlying data refreshes.
Knowledge base search can associate one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. When nothing is retrieved, the output path still completes. After retrieval output passes to the answering stage, the answer text and the cited source text need to be checked line by line. Hybrid retrieval merges multiple recall paths through RRF, reranking requires a rerank model that is enabled by configuration, and image retrieval requires corresponding image input and model capability. These paths do not all execute on every request. Semantic recall and relevance scoring are ranking and filtering mechanisms; they do not perform permission or entity isolation. Knowledge bases with different vector models cannot be associated together, so identifiers must be checked and non-matching bases excluded.
Constraints at the citation tracing step
The citation tracing step sits between retrieval and the operating report. The search node returns a citation array, and the downstream answering stage consumes it. A citation that cannot be matched to a source document, a version and a field cannot support a conclusion in an anomaly review. The relevance threshold filters low-similarity results only when semantic retrieval is used or when reranking is enabled, and the threshold is a filter, not a sort order. When question optimization, hybrid retrieval, image retrieval or reranking is active, the final result may be fused from several recall paths and will not necessarily follow a single vector similarity order. The citation limit is expressed in tokens per search, not in a fixed item count.
Hypothetical paper exercise sample: a media buyer's review material dated 10 June 2025 records that "Express plan A spent 8,000 yuan from 1 June to 7 June, and column 3 of the conversion attribution table shows 42 orders." The same material contains a draft sentence awaiting judgment: "This plan converted 42 orders from 1 June to 7 June, sourced from column 3 of the conversion attribution table." The check compares the draft against the source on subject, period and field. The material does not state a version number for the attribution table, so that field remains unknown.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only the campaign review base confirmed to share one vector model | Compare vector model identifiers across bases; exclude any base whose identifier differs |
| Search mode | Hybrid retrieval | Compare exact-term hits for plan names and identifiers against semantic-only retrieval |
| Citation limit | Start at a value covering the full sample material | Compare against the source text length; raise and retest if the source is truncated |
| Minimum relevance | Start at a value that retains the sample source text | Compare against the line-by-line check result; lower if the source text is filtered out |
| Result reranking | Enabled or not according to the actual configuration | Confirm the rerank model is available before enabling; compare against the result with reranking off |
| Citation output handoff | Check whether the citation array reaches the answering stage | Use a paper input of an empty array to check the downstream design |
Three common errors
- Treating semantic recall or a relevance score as a permission filter. Input: a cross-department query. Compare the recalled results against the set of materials the requester is authorized to access, and observe whether unauthorized material appears. A relevance setting does not perform entity isolation.
- Describing RRF as a toggle switch. Input: a hybrid retrieval configuration. Compare the documented description against the interface fields and observe whether a control named RRF exists. In the documentation, RRF is a fusion process applied to multiple result sets.
- Treating an unstated field as a contradiction. Input: the sample material. Compare the draft sentence against the source on each field and observe the missing field. An unstated field is pending verification, not a negation or a conflict.
How to verify the setup
- Input the sample source text and the draft sentence. Compare subject, period and field. The citation is traceable only when all three match the source; record any difference otherwise.
- Input a paper citation array
[]. Compare whether the downstream node still executes, and observe whether the output path completes as the documentation describes for empty results. Do not claim that a rule-based search must return an empty array. - Input two knowledge bases with different vector model identifiers. Compare the identifiers and observe whether the non-matching base is excluded. Record only the identifier check result; do not run a cross-model retrieval to compare missed recalls.