What the documents look like
On-demand retail store fulfillment processes online orders by the minute. This differs from periodic stocktaking, where counts are reconciled on a cycle. Operating reports and exception reviews for this category work order by order. The everyday materials are the instant order, the picking list, the out-of-stock report, and the delivery exception record, each updated at minute-level frequency per order.
The citation tracing task is to let every conclusion in a report return to the corresponding source document and state the effective scope of that source. Because materials update quickly, the version cited during a review may not match the version in force when the event occurred. Reconciliation must lock the document identifier and the time window.
A hypothetical drill sample is used throughout this page. The delivery exception record for one store, covering 12 March 14:00 to 15:00, states: "Order number A-0312-014, picking completed 14:22, rider arrived at store 14:41, customer signed 15:06." The out-of-stock report states: "Order number A-0312-014, item X out of stock, customer notified by phone at 14:18, customer agreed to substitute item Y." The draft sentence to be judged reads: "This order was delayed because of the stock shortage, and the store did not report status in time." Comparison must set the draft against both source texts. The sources record the shortage notification and the time points. They do not record the cause of any delay, and they do not record the time of, or absence of, a status report.
Constraints at the citation tracing step
The knowledge base search node outputs a citation array, and the length may be 0. When nothing is found, the output path still runs. After the citation array passes to the answering step, each source in the answer should map to cited source text. Retrieval is limited to the selected knowledge bases, and those knowledge bases must use the same vector model. Retrieval is not authorization. A similarity score does not perform entity isolation or permission filtering.
Full-text retrieval suits exact queries such as numbers and models. Semantic retrieval suits natural-language questions. Hybrid retrieval merges multiple result paths through RRF. Rerank runs when a rerank model is available and the setting is enabled. Image retrieval requires image input and the corresponding model capability. These optional paths in the documented flow are not executed on every request.
The relevance floor is a filter threshold, not a final ranking rule. It applies only in semantic retrieval or when result reranking is used. The citation limit is measured in tokens, not in a guessed item count.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to citation tracing for this page's sample.
| Configuration or check | Use on this page | Value or judgment method |
|---|---|---|
| Associated knowledge bases | Limit retrieval scope | Check the selected knowledge bases and material versions; confirm the same vector model is used |
| Search mode | Match document numbers | Use full-text or hybrid retrieval for order numbers and time points |
| Citation limit | Control cited content volume | Measured in tokens; a hypothetical trial starting value of 2000 tokens is chosen because the sample documents are short, compared with 4000 tokens to observe truncation |
| Relevance floor | Filter low-relevance results | Applies only in semantic retrieval or result reranking; a hypothetical trial starting value of 0.3 is compared with 0.5 to observe recall changes |
| Result reranking | Reorder text candidates | Requires a rerank model and must be enabled by setting; when not enabled, this path does not run |
| Citation content output | Carry empty results | Output is an array whose length may be 0; downstream design must accept an empty array |
Three common errors
- Treating the causal claim in the draft sentence as a source conclusion. The sources record the shortage notification and time points. They do not record the cause of any delay, so the causal judgment is unsupported.
- Treating "did not report status in time" as an established fact. The sources do not state the time of, or the absence of, a status report. This item remains to be verified and cannot be judged as a denial or a conflict.
- Treating retrieval output as authorization or an execution record. Retrieval returns cited content only. It does not complete approval, database writes, or external synchronization; those actions require separate system implementation.
How to verify the setup
- Input the two source texts and the draft sentence from the hypothetical sample. Compare whether the draft contains causal words or status judgments not present in the sources. If the draft keeps "because of" and "did not report in time," it fails. If it is rewritten to state only the shortage notification and the time points, it passes.
- Input order number A-0312-014. Compare whether the citation array from full-text or hybrid retrieval contains that document's source text. If the array is empty, record the gap and check whether the data has completed training. If it contains the document, then check each answer statement against the cited source text.
- Input an explicit paper empty array []. Compare whether the downstream answering step continues on an empty result. If the downstream step errors out and stops because of the empty array, the carrying design fails. If it continues and indicates no cited content, it passes.