What the documents look like
Marketplace advertising adjusts bids against live traffic and conversion, so campaign plans, keyword reports, audience package descriptions and conversion attribution tables refresh by day or by hour. Procurement and supplier records sit in a different document family. They describe the purchased object, the supplier entity, quoted prices and delivery stages. Citation tracing for this second family asks a narrow question: when a reply or handover note names a supplier, a price or an authorization state, does that statement correspond to a checkable source passage, and does that passage fall inside the currently valid document scope.
The supplied brief marks this category as roughly accurate rather than confirmed, so the descriptions below stay at the level of document behavior and search-node behavior. Update frequency matters only where it changes citation validity: a supplier quotation carries a validity window, and a dated campaign report does not. That difference affects which passage can support a handover sentence.
Constraints at the citation tracing step
The knowledge base search node outputs cited content as an array, and the array length can be zero. A zero-length output means the downstream branch still executes; it does not mean the workflow terminated with an error. Before retrieval, the selected knowledge bases and document versions need checking, and only knowledge bases sharing the same vector model can be associated together. The minimum relevance value is a filter threshold for semantic retrieval or result reranking; it does not perform permission isolation or entity separation. Hybrid retrieval merges multiple recall paths through RRF, and reranking requires a rerank model that is enabled in the configuration. Whether retrieval output is carried into the answering step, and whether each answer sentence corresponds to a cited source passage, remain manual checks.
The documented search modes behave differently on exact strings. Full-text retrieval suits numbers, model codes, proper nouns and error codes, while semantic retrieval suits natural-language questions. Hybrid retrieval combines both and merges through RRF. The minimum relevance setting takes effect only under semantic retrieval or when result reranking is used, and the citation limit caps the tokens of cited content per search rather than a fixed item count.
A hypothetical exercise sample is used below. It is labelled as hypothetical and does not describe any real supplier, real company or real authorization record.
Hypothetical source text, from a supplier library document titled Annual Framework Procurement Intent Record:
"Supplier entity: a packaging materials company in East China; object: corrugated cartons; stage: quotation submitted, awaiting purchaser review; quotation validity through 31 March 2025; authorization state: purchase contract not signed."
Hypothetical draft handover sentence awaiting review: "This supplier has passed review, and carton supply runs through 31 March 2025."
The question is whether the draft sentence can be supported by the source text above.
Configuration and implementation checks
| Configuration or check item | Hypothetical exercise starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the supplier library | Check the selected library identifier and vector model identifier, and exclude libraries on a different vector model |
| Search mode | Hybrid retrieval | Supplier names and codes need exact hits while semantically close passages also need recall |
| Minimum relevance | Set to 0.3 | A starting point for this exercise, compared against the number of recalled passages when no threshold is set |
| Citation limit | Set to 2000 tokens | A starting point for this exercise, compared against the displayed passage count before adjustment |
| Result reranking | Enabled | When candidates are numerous, filter by rerank score and compare against ordering by vector score alone |
| Citation output handling | Manual review | Check whether each answer sentence shares the same entity, field and period as the cited source passage |
These values are exercise starting points, not documented defaults. The search node itself does not approve suppliers, sign contracts or write to any procurement system. Those actions require separate downstream implementation that is not described in the supplied documents.
Three common errors
- Reading "quotation submitted, awaiting purchaser review" as "has passed review." The review status field holds the value "awaiting purchaser review," and the draft sentence writes "has passed review." These are mutually exclusive values in the same field, so the draft sentence is unsupported.
- Treating "quotation validity through 31 March 2025" as a supply deadline. Quotation validity and supply deadline are two separate fields. The source text does not state a supply cut-off, so the draft phrase "supply runs through 31 March 2025" has no source basis.
- Treating "purchase contract not signed" as an authorization that took effect or ended. Not signed describes the signing fact. Whether the contract became effective and whether cooperation was reached are separate facts requiring separate verification, and neither follows from this field.
How to verify the setup
- Enter the Annual Framework Procurement Intent Record source text and the draft sentence above, then compare the review status field. The source holds "awaiting purchaser review" and the draft holds "has passed review." Because the two values are mutually exclusive within the same field, the correct observation is a contradiction, and the draft sentence is marked unsupported.
- Enter the same source text and the same draft sentence, then compare the supply deadline field. The source states only "quotation validity through 31 March 2025" and states no supply deadline. The correct observation is that the claim is unsupported, and the evidence gap is recorded rather than filled.
- Enter the same source text together with the "purchase contract not signed" field, then compare three separate facts: signing, contract effectiveness and cooperation reached. The source supports only the not-signed fact. The other two remain unverified, and the result must not be recorded as authorized or as terminated.
For the empty-output branch, a paper input of an empty citation array can be used to check whether the downstream design still proceeds. This checks the documented behavior that a zero-length citation output keeps the branch running; it does not assert that any particular query returns an empty array. When a library uses a different vector model, the check is to compare identifiers and exclude the mismatched library from association, without running a cross-model retrieval to compare recall.