What the documents look like
Marketplace advertising adjusts bids against traffic and conversion in near real time, and platform campaign review is not the governing gate. The working materials are campaign plans, keyword reports, audience package descriptions and conversion attribution tables. These materials can update daily or hourly, so the same campaign plan may carry different bid, budget and attribution wording on different dates. Contract and transaction material comparison has to map a proposed sentence back to source text line by line. At the citation tracing step this means confirming the material version and field scope first, then deciding whether a drafted sentence or handover request has matching source support.
A campaign plan may state a daily budget ceiling, a flight period and a channel. A keyword report or attribution table may state a reporting deadline and a reporting object. Those are the fields a comparison can actually check. Anything the source text does not state stays unstated, and the comparison records a gap rather than filling it.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model for vector search. It outputs cited content as an array, and the length may be zero; when nothing is found, that output path still runs through. Hybrid retrieval uses semantic recall and full-text recall together and merges the two result sets through RRF. A ReRank model reorders candidate text and returns a 0-1 score. Minimum relevance takes effect only in semantic retrieval or when result reordering is used. The citation cap controls the maximum amount of content cited per search, measured in tokens, and it is unrelated to any thinking budget.
These mechanisms serve retrieval and ordering. They do not perform permission or entity isolation, and a retrieval result is not an authorization to act. A relevance score is not a category, entity or permission filter, so it cannot be used to conclude that a business is unrelated or a record is unauthorized.
Hypothetical paper exercise sample. Source text one: "Party A commissions Party B to place advertising on Juliang Qianchuan, with a daily budget ceiling of RMB 8,000, for the flight period 1 March 2025 to 31 March 2025." Source text two: "Party B shall submit the previous calendar month's keyword report and conversion attribution table before the fifth day of each month." Draft sentence to be judged: "Party B is authorized to raise the daily budget to RMB 12,000 in March and to submit the March report before 10 April." The budget change and the submission date both need to be checked against the source text. The source text does not state any budget change authority, and it does not state a 10 April submission arrangement.
Configuration and implementation checks
The table below concerns this page's sample. It lists documented settings or implementation checks only. Where a number is needed, it is marked as a hypothetical starting point for this exercise.
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases sharing the same vector model | Only the same vector model can be associated together; verify the identifier first and exclude non-matching ones |
| Search mode | Hybrid retrieval | Uses semantic recall and full-text recall together, merged through RRF |
| Result reordering | Enabled | Reorders candidate text; the 0-1 score can be used for filtering |
| Minimum relevance | Hypothetical starting point 0.5 | Compared against the reorder score; effective only in semantic retrieval or with reordering |
| Citation cap | Hypothetical starting point 3000 tokens | Compared against the length of returned cited content; trimming follows tokens |
| Cited content output | Observe the array length | Length may be zero; an empty array does not mean the workflow errored out |
Three common errors
- Treating retrieval output as authorization to act. Input: the cited array returned by knowledge base search. Comparison: the draft sentence's claim that budget increase is authorized. Observable result: the cited source text does not state any budget change authority, so that sentence is unsupported.
- Associating knowledge bases that use different vector models. Input: two knowledge bases with different vector model identifiers. Comparison: the node's requirement that associated knowledge bases share the same vector model. Observable result: the identifiers do not match, so the non-matching one is excluded, and no cross-model retrieval comparison is run.
- Treating a relevance score as a permission or entity filter. Input: one cited item with a low relevance score. Comparison: the draft sentence against the source text's subject, period and fields. Observable result: the score affects ordering and filtering only; it cannot be used to decide that a subject lacks authority or that a record is isolated.
How to verify the setup
- Check the budget field against source text one and the draft sentence. Input: source text "daily budget ceiling of RMB 8,000" and draft sentence "raise the daily budget to RMB 12,000". Comparison: the value of the same field. Observable result: the two values are mutually exclusive, the draft sentence is unsupported by the source text, and the evidence gap is recorded.
- Check the submission date against source text two and the draft sentence. Input: source text "before the fifth day of each month" and draft sentence "before 10 April". Comparison: the date value of the same field. Observable result: the source text does not state a 10 April arrangement, so that sentence remains unverified.
- Check downstream handling with an explicit paper input of an empty array. Input: a cited content array of length zero. Comparison: whether the downstream node still continues execution. Observable result: the path runs through and produces no cited source text; this does not support any claim that a retrieval rule necessarily returns an empty array.
The comparison must stay on the same subject, the same period and the same field. Only an explicit mutually exclusive value in the same field counts as a contradiction. Text that is unstated, or that only shows a related stage, counts as unsupported or pending verification, not as a denial or a conflict. A pending review does not show that an operating permission is absent. Signing, reaching cooperation and contract effectiveness are three separate facts. A quotation validity period does not indicate a supply period. No authorization or effectiveness clause is added to force a definite answer, and the evidence gap is kept open.