What the documents look like
Franchise filing management handles legal disclosure and the commerce department system. It does not handle individual store siting or construction. For the procurement and supplier records task, the working material consists of contract ledgers, store lists, enterprise licences and filing submissions. Citation tracing means each cited statement can be returned to a specific document and version.
Updates are irregular or annual. A supplier record may therefore exist in a historical version and a current version at the same time. Filing errors can lead to an order to file, franchise contract disputes or administrative penalties, so the point of citation tracing is to let every citation reach its source text.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Its output is an array of citations, and the length may be zero. The citation content is passed to the answering step, where each item is checked against the cited source text. Semantic recall and relevance scores affect ordering and filtering; they do not carry permission or entity isolation. The same-vector-model requirement applies to knowledge base association only. It does not extend to requiring the same model for a commerce website or any other external system.
Hypothetical paper exercise sample: the source text to be checked reads, "The supplier list submitted by Franchisor A within the filing year states that Supplier B has operating status, and the effective scope is the filing materials for that year." A draft sentence to be judged reads, "Supplier B's supply period covers the following year." The source text states only the filing material scope for that year and does not state a supply period for the following year, so the draft sentence is not supported. If the draft sentence is changed to "Supplier B has operating status in the filing materials for that year," it matches the source text on the same subject, the same period and the same field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Filing materials base only | Check the vector model identifier before associating; exclude bases that do not match |
| Search mode | Hybrid search | Compare against semantic search and full-text search separately, observing whether citations reach the source text |
| Minimum relevance | 0.5 as the hypothetical starting point | Compare against 0.3 and 0.7, observing whether unrelated fragments decrease |
| Citation limit | 2000 tokens as the hypothetical starting point | Compare against 1000 and 4000, observing whether citations are truncated |
| Result reranking | Enabled as configured | Compare against not enabled, observing whether ordering stays closer to the source text |
| Question optimization | Enabled as configured | Compare against not enabled, observing whether follow-up questions reach the source text |
These values are hypothetical starting points for this exercise, not official defaults. The citation limit uses the token unit, which differs from top k. RRF is a fusion process for mixed results and is not a switch.
Three common errors
- Treating a semantic recall score as a permission or entity filter. Check the input as the same query under different authorization scopes, compare whether out-of-scope source text appears, and observe whether the result matches the authorization scope.
- Writing a pending-review status as signed or effective. Check the input as the sample source text and the draft sentence, compare subject, period and field, and observe whether the draft sentence exceeds the source text scope.
- Treating an empty citation array as a workflow error. Check the input as the paper input
[], compare whether the downstream node continues, and observe whether the output path completes.
How to verify the setup
- Input the sample source text and the draft sentence "Supplier B's supply period covers the following year." Compare whether the source text effective scope states the following year. Observe whether the draft sentence is marked as not supported.
- Input the sample source text and the draft sentence "Supplier B has operating status in the filing materials for that year." Compare whether subject, period and field match. Observe whether the draft sentence matches the source text.
- Input the paper citation array
[]. Compare whether the downstream node accepts it. Observe whether the output path completes, without claiming that a fictional rule search necessarily returns[].
Knowledge bases with different vector models cannot be associated together. Check the identifier first, exclude those that do not match, and record the actual observation. Missing evidence remains a gap. Pending review does not demonstrate a lack of operating permission.