What the documents look like
Platform advertising adjusts bids against traffic and conversion in real time, so the working set keeps moving: delivery plans, keyword reports, audience package descriptions and conversion attribution tables update on daily and hourly rhythms. Multilingual and cross-border retrieval faces the same set of documents in several language versions, and the hard part of citation tracing is confirming which document, which version and which passage a given citation actually points to. Delivery specialists, brand marketing teams and agency operators reconciling attribution need to line up retrieval output against the original delivery documents item by item. Wasted ad spend and distorted ROI follow directly when the source of a citation stays unclear.
Constraints at the citation tracing step
The knowledge base search node recalls content from already linked knowledge bases and outputs citations as an array, whose length may be zero. A zero-length citation output means this run returned no entries; it does not mean the workflow errored out or stopped, and it does not mean the material does not exist. Semantic retrieval and the minimum relevance value govern recall and filtering only, and carry no permission or entity isolation: retrieving an entry does not mean the caller holds authorization to view or use it. Linked knowledge bases must share the same vector model; bases on different vector models cannot be linked together. At the check step, confirm the identifiers first and exclude those that do not match, record what is actually observed, and do not run cross-model retrieval to compare missed recalls.
Hypothetical paper exercise sample: the source material to be checked consists of three sentences. Sentence one: "This delivery plan applies from March 1, 2025 to March 31, 2025, to self-operated stores in the East China region." Sentence two: "In this plan's keyword report, the brand-term bid ceiling is 2.5 yuan per click." Sentence three: "This plan's audience package description is marked as draft status and has not been submitted for review." The draft sentence to be judged is: "This delivery plan has passed review, and the brand-term bid may be executed." The judgment compares the same subject, the same period and the same field: the material states only that the audience package description sits in draft status and records no review approval, so the draft sentence's claim of passed review is unsupported, and the claim that the bid may be executed likewise lacks a basis.
Configuration and implementation checks
| Check item | Relevant setting or action on this page | Basis for judgment |
|---|---|---|
| Linked knowledge bases | The knowledge base search node selects already linked knowledge bases | Check vector model identifiers, exclude those that do not match, do not run cross-model retrieval |
| Citation output | The search node outputs citations as an array, length may be zero | Treat an empty array as no returned entries, do not presume an errored stop |
| Citation input handling | Check whether retrieval output is passed to the answering step | Compare answers against cited source text item by item, do not substitute fluency for source evidence |
| Citation display check | Use a traceable sample to check displayed source text | Check whether subject, object, period and status in the source match the citation |
| Material timeliness check | Test sources with current and historical document versions | Whether a citation is adopted still depends on review against the effective version, retrieval alone does not establish effect |
| Hypothetical trial starting value | This hypothetical trial sets the minimum relevance starting value at 0.5 | Chosen to make entry differences before and after filtering observable, compared against hypothetical values 0.3 and 0.7 for changes in recalled entry counts |
Three common errors
- Treating an empty citation array as a retrieval failure. The input is one citation output returned with length zero; the comparison object is whether the workflow errored out and stopped; the observed result is that the chain continues and outputs an empty array. Record this as no returned entries, and do not write it up as a system fault.
- Treating a retrieval hit as authorization. The input is a delivery plan entry returned by retrieval; the comparison object is whether the caller holds viewing or usage permission; the observed result is that the retrieval action itself produces no authorization, and permission must come from another source. Record this as pending verification.
- Reading draft status as passed review. The input is sentence three of the hypothetical sample, the audience package description marked as draft; the comparison object is the draft sentence's claim of passed review; the observed result is that the material records no review approval. Judge this as unsupported.
How to verify the setup
- Input the three source sentences of the hypothetical sample and the draft sentence; compare whether the draft sentence's claimed passed review and the source's recorded draft status share the same subject, the same period and the same field; the observed result is that the draft sentence is unsupported. When that condition holds, record the judgment, and do not fabricate authorization or effective clauses.
- Input a paper citation array [] and compare how the downstream design handles an empty array; the observed result is that downstream treats it as no returned entries and the chain does not error out. When that condition holds, confirm the empty-result handling, and do not claim that a fictional rule search must return an empty array.
- Input the list of vector model identifiers for the already linked knowledge bases and compare whether the identifiers match; the observed result is that non-matching ones are excluded and no cross-model retrieval is run. When that condition holds, confirm the linked-scope check is complete, and do not invent runtime results after joint linking.