What the documents look like
Convenience-store fresh food is ordered by the store at high frequency against short shelf lives. Each day or each shift produces an order sheet, a loss report, a near-expiry handling sheet and a temperature record. These four materials are the working set. POS sales, the headquarters ordering system, store inventory and cold-chain temperature records each capture a different fact, so a missing item removes the basis for an answer. Citation tracing requires that every statement in an after-sales reply map back to a checkable line in the original text. Fluent wording does not replace source evidence.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. An empty array still lets the output path run through; it does not mean the workflow has failed and stopped. Search results must be handed to the answering step, and each answer statement must be checked line by line against the cited original. A similarity value only filters low-relevance content; it does not perform permission or entity isolation. Retrieval is not authorization, and a form collecting values is not the same as business verification, approval, a database write or an external synchronization having been completed.
Hypothetical drill sample: a store submits a fresh-food order sheet on the morning shift of 10 March 2025. The original text reads "Order date 2025-03-10, item A, quantity 12, store submitted." The same day's temperature record reads "03-10 morning shift, refrigerated cabinet 4 degrees Celsius, recorder signed." The draft sentence to be judged is "This order sheet was approved on 2025-03-10 and delivery can be arranged." The original text only states that the store submitted; it does not record an approving party, an approval status or delivery authorization. The draft sentence is unsupported, and the approval evidence gap must remain open.
Configuration and implementation checks
| Configuration or check item | Hypothetical trial starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Fresh-food ordering document base | Check the selected base identifier against the vector model identifier; exclude bases whose vector models differ |
| Search mode | Hybrid search | Item-code questions need full-text hits; natural-language questions need semantic recall |
| Citation limit | 800 tokens as the hypothetical trial starting point | Compare against the length of a single order sheet or temperature record and observe whether citations are truncated |
| Minimum relevance | 0.4 as the hypothetical trial starting point | Compare against the score distribution of historical samples and observe whether low-relevance fragments are filtered |
| Question optimization | Enabled | Continuous follow-up needs coreference resolution; compare retrieval output with the disabled state |
| Result rerank | Enabled according to settings | Reorder when candidates are numerous; compare ranking change against the unreranked result |
These values are hypothetical trial starting points for this page and are not official defaults. The table lists only settings or implementation checks tied to this page's feature.
Three common errors
- Treating a retrieval hit as authorization. Input the order sheet original text above together with the draft sentence, compare the fields "submitted" and "approved", and the error appears when the answer outputs an approval conclusion directly. The observable result is that the draft sentence is unsupported by the original text.
- Treating an empty citation array as a system fault. Input a paper citation output of [] and compare whether the downstream design continues to run; if the downstream is configured to error out and stop, that conflicts with the documented behavior.
- Associating knowledge bases that use different vector models. Input two base identifiers, compare whether the vector model identifiers match, and if they do not match while association is still attempted, the association condition is violated. Do not run a cross-model retrieval to compare missed recall.
How to verify the setup
- Citation-to-original check. Input the order sheet and temperature record originals above, and compare whether each citation in the answer can be found as a corresponding sentence in the original. The condition is met when the cited content agrees with the original on subject, period and field.
- Empty-result handling check. Input a paper citation output of [] and compare whether the downstream node continues to execute. The condition is met when the downstream does not error out and records the gap.
- Document-currency check. Input the current version and a historical version of the order sheet, and compare which version the citation source points to. The condition is met when the answer marks the version and does not treat the historical version as currently effective. These are expected check methods and no actual test is claimed.