What the documents look like
Marketplace advertising adjusts bids continuously against traffic and conversion, so its working documents change on a daily or per-order rhythm rather than waiting on a platform campaign review. The records this page concerns are the delivery plan, the keyword report, the audience package description, and the conversion attribution table. A delivery specialist reads them daily or hourly, and each revision can change whether an answer about an advertising order, a refund, or a material handover still points to a valid source.
For order fulfilment and after-sales replies, the practical difficulty is version drift. A refund answer traced to an earlier version of the delivery plan may describe a window that the current version has already replaced. Citation tracing therefore has to name the exact document version behind each statement, and it has to keep the answer separate from the source it cites.
Constraints at the citation tracing step
The knowledge base search node accepts one or more knowledge bases that share the same vector model. Its output is an array of citations, and that array may have length zero, which means the downstream chain still runs when nothing is found. A citation array is retrieval output, not an authorization record. A relevance score is a similarity value, not a category filter, an entity filter, or a permission filter, and it does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization. Forms collect values; business verification, approval, database writes, and external synchronization each require their own stated downstream implementation.
Hypothetical paper exercise. The material text under review reads: "The 1 March 2025 edition of the Spring Delivery Plan states: this plan applies to advertising delivery orders placed between 1 March 2025 and 31 March 2025, and a refund request must be submitted within 7 days after order payment." The draft sentence to be judged reads: "The refund request for this order has been approved." The material names a subject, an object, and an effective range, but it records only a submission deadline. It does not record an approval field or an approval status. The draft sentence is therefore unsupported, and the approval state remains unknown. A pending request is not approval, and an offered option is not execution.
Configuration and implementation checks
| Setting or check | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Two knowledge bases sharing the same vector model | The document permits joint association only under the same vector model; verify the identifiers first |
| Citation limit | 2000 tokens as a hypothetical starting value | The document controls citations by tokens; compare against the length of the sample material |
| Minimum relevance | 0.5 as a hypothetical starting value | The document states this value takes effect only in semantic retrieval or with reranking |
| Search mode | Hybrid retrieval | The document states vector and full-text retrieval run together and merge through RRF |
| Citation output | Check whether the output array reaches the answering step | The document states length may be zero; an empty array is not a workflow error |
| Material version check | Test with the current and a historical document version | The document requires combining the citation with the effective version review |
Hybrid retrieval and RRF describe a merge process, not a switch to be toggled. Image retrieval requires an actual image input and matching model capability, and reranking requires a rerank model enabled under the settings. The optional paths in the overall retrieval diagram do not all execute on every request. The same-vector-model requirement applies to associated knowledge bases; it does not extend to business registration sites or other external systems.
Three common errors
- Treating retrieval output as authorization to act. The input is the citation array returned by the search node; the comparison is whether that array reaches the answering step and whether each answer line matches the cited original text; the observable result is an answer that corresponds line by line to the source. Retrieval by itself is not authorization.
- Treating a relevance score as a permission filter. The input is the minimum relevance value of 0.5; the comparison is whether returned content includes unrelated business records; the observable result is whether content outside the intended scope appears. Relevance does not exclude unrelated businesses.
- Writing a pending state as approved. The input is the draft sentence "The refund request for this order has been approved"; the comparison is whether the material records an approval field; the observable result is that the material records only a submission deadline. A pending request does not establish approval.
How to verify the setup
- Input the hypothetical material text and the draft sentence given above; compare whether the material records an approval status; observe whether the draft sentence is unsupported. The condition is met when the material has no approval field and the sentence is judged unsupported.
- Input an empty citation array shown as []; compare whether the downstream answering step accepts it; observe whether an error interrupts the chain. The condition is met when the downstream design handles a citation length of zero. This is a paper check of downstream design, and it does not claim that any real query returns an empty array.
- Input two knowledge base identifiers that use different vector models; compare whether the identifiers match; observe whether the non-matching one is excluded. The condition is met when only knowledge bases with the same vector model are associated. Do not run a cross-model retrieval comparison to measure missed recall; record the identifier check and the actual observation instead.
For each check, the evidence needed must sit inside the accessible target retrieval scope or inside an explicit manual input. When it is absent, record the gap rather than filling it. Thinking budget is a reasoning budget and does not serve as a citation token limit. An unsigned termination is not a completed termination, and a quotation validity period does not state a supply period.