What the documents look like
A fresh-food concession counter in a hypermarket is operated by an on-site supplier, with revenue shared by sales. Order fulfilment and after-sales answers therefore involve the hypermarket fresh-food department, the concession supplier and the store manager at the same time. Daily materials include the concession contract, daily sales reports, loss sheets and food incoming inspection records. Data sources cover the POS system, supplier delivery notes, food-safety ledgers and store stock counts. These materials update daily or per batch. Reconciling weighed sales, handling loss allocation and checking incoming goods certificates are the most time-consuming steps.
Citation tracing answers one question: when replying to a customer or handing over internally, can each quoted passage be matched item by item to a specific document and a valid version? The contract fixes who bears loss and how it is shared; the inspection record fixes whether a batch was checked. A fluent answer that merges these into one claim loses the ability to show which document supports which part.
Constraints at the citation tracing step
Knowledge-base search can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. At this step, check the vector model identifier of each candidate base and exclude those that do not match, then record what was actually observed. The search node outputs cited content as an array, and its length can be 0. An empty array does not mean the workflow has errored or stopped; downstream design must accept an empty result.
Semantic retrieval, full-text retrieval, hybrid retrieval and result re-ranking are selectable paths. Re-ranking requires a re-rank model and must be enabled in settings. Image retrieval requires image input and the corresponding model capability. The minimum relevance value applies only under semantic retrieval or when result re-ranking is used. It is a filtering threshold and does not perform permission or entity isolation. After the retrieval output is passed to the answering step, the answer and the cited source text must be checked line by line.
Hypothetical paper exercise sample: the concession contract excerpt reads, "Party A is the fresh-food department of XX hypermarket; Party B is the on-site supplier; loss sharing is reconciled batch by batch against the daily sales report and the loss sheet, for the period 1 March 2025 to 31 March 2025." The food incoming inspection record excerpt reads, "12 March 2025, batch A-0312, delivery note SC-0312, inspector has signed, status: inspected." A draft sentence to be judged reads, "This batch has completed inspection and loss-sharing settlement has been completed under the contract." Judgement point: the inspection record supports "inspected"; loss-sharing settlement is not stated in the given text, so the second half of the sentence is unsupported.
Configuration and implementation checks
| Check item | Hypothetical starting point for this exercise and reason | Comparison method and observation |
|---|---|---|
| Associated knowledge bases | Select the four document bases: concession contract, daily sales report, loss sheet, food incoming inspection record | Compare the vector model identifier of each base; exclude any that differ |
| Citation cap | Start at a size covering one contract clause plus one inspection record, because the material checked in a single pass here is short | Compare against the actual length of cited content; observe whether truncation occurs |
| Minimum relevance | Start at a stricter value, because after-sales answers must trace back to source text | Compare against the re-rank score of returned citations; observe whether filtering occurs |
| Search mode | Hybrid retrieval, because batch, number and note-number queries depend on full-text recall | Compare hit differences between number queries and natural-language queries |
| Cited content output | Accept the array output and allow length 0 | Use a paper input of [] to check whether the downstream path runs through |
| Answering step handover | After retrieval output is passed to the answer node, check line by line | Compare the correspondence between answer sentences and cited source text |
Three common errors
- Forcing knowledge bases with different vector models into one association. Input: two bases whose vector model identifiers differ. Compare the two identifiers. The non-matching base should be excluded; record the observed retrieval result as it is, without inferring missed recall or redundancy.
- Treating an empty citation array as a process failure. Input: a paper []. Compare whether the downstream node continues to execute. The condition is met when the downstream path runs through and the empty result is recorded.
- Expanding "inspected" into "loss-sharing settlement completed." Input: the inspection record and contract excerpt above. Compare the draft sentence with the source text. The condition is met when the draft keeps only the part supported by the source and marks the rest as unsupported.
How to verify the setup
- Input the contract excerpt and inspection record text above, then compare whether the answer citations correspond item by item to those two materials. The condition is met when each citation can be traced back to a specific document and date range.
- Input the draft sentence "This batch has completed inspection and loss-sharing settlement has been completed under the contract," then compare it with the source text. The condition is met when the judgement output separates "inspected" from the unstated settlement and preserves the evidence gap.
- Input a paper [] to check downstream handling, then compare whether the node continues to execute. The condition is met when the flow runs through and the empty result is recorded. This does not claim that a real retrieval will necessarily return [].
These checks concern the citation tracing step only. Whether a handover material, an operation or a request is authorized at a given point is a separate matter; retrieval and form entry do not by themselves constitute authorization to act.