What the documents look like
Ordinary platform refunds and returns arise order by order, and they do not necessarily attach to a promise made in a single livestream session. Citation tracing for procurement and supplier material therefore has to draw on several record streams at once: the platform after-sales console, customer service chat, WMS receiving records, and payment refund records. The working documents are refund requests, return logistics notes, quality inspection photos, and refund approvals. Within procurement and supplier material, the three most time-consuming steps are deciding return responsibility, chasing the returned parcel, and approving the refund amount. Each of these steps produces records that a later reader may need to trace back to an original source.
For citation tracing, the practical consequence is that a fluent answer is not evidence of origin. A refund request may cite a chat message; a quality inspection photo may support or fail to support a damage claim; a refund approval may rest on a WMS receiving record. The tracing step asks whether each statement in a draft can be matched to a specific source document, and whether that source actually carries the fact being asserted. Where the source is silent, the statement remains unsupported. Where the source states a different value for the same field, the statement is contradicted. These two outcomes are not the same, and neither is the same as an approval or a completed termination.
Constraints at the citation tracing step
The knowledge base search node outputs citations as an array, and the length may be zero. A zero-length output still lets the downstream chain run; an empty array is not by itself a workflow failure. Citation tracing must confirm whether the search output is passed to the answering step, and then compare the answer against the cited original text item by item. Material currency checks use a current version and a historical version of a document as test sources; whether a citation is adopted still depends on review against the effective version. Semantic recall and relevance scores do not perform permission or entity isolation. Knowledge bases may be associated together only when they share the same vector model, and that requirement does not extend to business registration sites or other external systems.
Hypothetical drill sample: the source text reads, "Supplier A submitted a quotation on 1 March 2025, valid until 30 April 2025, for consumable item A used in return quality inspection, quantity 200 units." The draft sentence to be assessed reads, "Supplier A has confirmed supply, and the supply period covers May 2025." The draft treats the quotation validity period as the supply period and states no signing or contract effectiveness fact. The draft sentence is unsupported.
The distinction matters at three points. A quotation validity end date is not a supply period. A pending review is not a grant of operating authority. Signing, reaching cooperation, and contract effectiveness are three separate facts. Where the source does not state a term, the gap stays open rather than being filled by inference.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the after-sales material base whose vector model identifier matches | Compare against a base with a different vector model identifier; exclude the mismatched base |
| Citation upper limit | Set a hypothetical starting value sufficient to cover the sample source text in tokens | Compare against the sample source length to confirm the full text enters the citation |
| Minimum relevance | Set a hypothetical starting value of 0.5 | Compare against results with filtering off, and observe whether the sample source is retained |
| Search mode | Hybrid search | Compare against semantic search alone; the sample contains a number and proper nouns |
| Question optimization | Off | Compare against the enabled state; the sample question already has a complete subject |
| Citation output handling | Check that the answering step maps to cited original text item by item | Compare against reading the answer text alone, to confirm the source can be traced |
The citation upper limit is expressed in tokens, not in a top-k item count. The minimum relevance value applies only under semantic search or when result reranking is used, and it is a filter threshold rather than a final ordering rule. RRF is a fusion process for combining multiple retrieval channels; it is not a switch to be turned on. Image retrieval requires an image input and matching model capability, and reranking requires a rerank model enabled by configuration. No scheduled trigger or hidden control is assumed.
Three common errors
- Treating the quotation validity period as the supply period. Input: the sample source text and the draft sentence. Compare the quotation validity field against the supply period field. Observe whether the draft exceeds what the source states. Expected: the draft sentence is unsupported.
- Treating an empty citation array as a workflow error. Input: a paper output of []. Compare the node output length against the downstream handling design. Observe whether the downstream chain continues on a zero-length citation. Expected: no error is declared on the basis of the empty array.
- Treating semantic relevance as a permission filter. Input: an unauthorized record and a query. Compare the retrieval scope against the authorization boundary. Observe whether retrieval returns the unauthorized record. Expected: relevance does not perform permission isolation, and authorization must be checked separately.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the quotation validity field against the supply period field. Observe whether the draft exceeds the source range. The condition is met when the draft sentence is marked unsupported.
- Input a paper output of []. Compare the node output length against the downstream design. Observe whether the downstream chain continues. The condition is met when the downstream runs through on a zero-length citation.
- Input two knowledge bases with different vector model identifiers. Compare whether the identifiers match. Observe whether the mismatched base is excluded. The condition is met when only the matching base is associated and the actual observation is recorded.
These checks describe what to compare and what remains unknown. They do not assert that a first result will be returned, that all relevant items will be recalled, or that an unrelated query will return nothing. Where evidence is missing, it stays missing.