What the documents look like
Refund-only disputes frequently reach a decision without any returned goods arriving at a warehouse. The evidence therefore centres on dispatch and signed receipt, together with the value of the item. Procurement and supplier records become the reference material that support staff check against the dispute file.
That material covers purchase orders, supplier quotations, inbound and outbound warehouse records, dispatch documents and proof of receipt. The brief describes these records as updating per order and daily. This page does not depend on that cadence, so it is noted only as stated and no batch, full-sync, real-time or transaction-triggered mechanism is inferred from it.
The records mix two shapes of content. Purchase order numbers, item model codes and document numbers are precise fields. Supplier names and product descriptions are natural-language content. A single retrieval mode covers both shapes poorly, which is the practical reason the search mode choice matters for this category.
A second feature of this category is that a retrieved record is not an authorised record. A supplier quotation that surfaces in a search result says nothing about whether it may be shown in an external appeal or disclosed outside the team. Authorisation state is a separate question that retrieval does not answer.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. Checking library identifiers and excluding those that do not match is a manual step, and no cross-model retrieval is run to compare recall.
Search results are returned as an array of cited content, and that array may have a length of zero. The downstream path still runs when nothing is found. This is documented behaviour of the node, not a guarantee about any particular query.
Semantic retrieval, full-text retrieval and hybrid retrieval each have their own conditions. Hybrid retrieval combines the two paths through RRF, and Rerank performs a second ordering of candidate text. The minimum relevance value applies only under semantic retrieval or when result reranking is in use. It is a filter threshold, not a final ordering rule, and it is not a category filter, an entity filter or a permission filter. A similarity setting does not exclude unrelated businesses or unauthorised records.
The optional paths in the documented flow are not executed on every request. Image retrieval requires an image input and the relevant model capability. Reranking requires a rerank model and must be enabled under the settings. RRF is a fusion process applied to results from multiple paths, not a switch that is turned on.
Retrieval is not authorisation. A pending review request is not an approval. An offered option is not an execution. An unsigned termination is not a completed termination. A quotation validity period is not a supply period.
Hypothetical paper exercise. A support specialist needs to check supplier material against a refund-only dispute. The source text reads: "Supplier A issued a quotation on 10 March 2026, valid until 9 April 2026, item model A-100." The draft sentence under review reads: "Supplier A still supplied at the original quotation on 20 April 2026." The source states only a quotation validity end date. It does not state that any supply occurred, nor at what price, and 20 April 2026 falls after the stated validity end date. The draft sentence is unsupported. Whether supply occurred and on what terms remains unknown.
Configuration and implementation checks
| Item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the procurement and supplier records library | That library and the dispute documents use the same vector model; check library identifiers one by one and exclude non-matching ones |
| Search mode | Begin testing with hybrid retrieval | The records contain model codes and descriptive text; compare against pure semantic and pure full-text results |
| Citation limit | Start at 2000 tokens | Test with records of different lengths to see whether content stays complete; compare with 1000 and 3000 |
| Minimum relevance | Start at 0.4 | First label relevant and irrelevant samples, then compare filtering results at 0.3 and 0.5 |
| Question optimisation | Enable only for multi-turn follow-ups | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enable where a rerank model is available | Compare candidate ordering before and after reranking and observe the effect of score filtering |
These starting values are choices for this exercise. They are not documented defaults and no official limit, unit or global switch is claimed.
Three common errors
- Treating a quotation validity period as a supply period. The draft sentence asserts supply at the original quotation on 20 April 2026. The source records only that the quotation was valid until 9 April 2026. Supply and price are not recorded, so the assertion is unsupported.
- Treating a retrieval hit as authorisation. Finding a supplier record in a search result does not mean the record may be used in an external appeal or disclosed. Authorisation state must be checked separately.
- Treating a pending review as approved. When a handover request is pending review, that state does not establish that operating permission exists or that a business action has been carried out.
How to verify the setup
- Enter the source text and the draft sentence from the paper exercise. Compare the quotation validity end date with the date stated in the draft sentence. If the draft date falls after the validity end date and no supply is recorded, the draft sentence is unsupported. This is the expected result; it has not been run.
- Enter a query containing only a supplier name. Compare the returned cited content against the scope of the procurement and supplier records library. The link configuration meets the requirement when all returned content belongs to that library and no cross-model linking occurred. If returned content falls outside that library, or if libraries with different vector models were linked, the configuration does not meet the requirement.
- Enter an explicit paper input of an empty array. Observe the downstream node behaviour. The empty-result handling meets the documented description when the downstream path continues on empty cited content without error. This checks the downstream design against the documented statement that the output array may have length zero. It does not assert that any particular query returns an empty array.
Where the source text does not record a field, that field remains unknown. Missing evidence stays missing, and no approval, refund, deployment or other business action is claimed to have occurred.