What the documents look like
Fruit procurement runs on wholesale market quotations, supplier offers, store sales, and inspection reports. The daily working documents are purchase orders, acceptance records, pesticide-residue test reports, and loss records. These materials are updated daily or per batch, and acceptance grading together with loss control consumes a large share of staff time. Operating reviews and exception reviews must place materials from different batches, different stores, and different dates inside one retrieval scope so that the same field can be compared across documents.
A frequent difficulty is that the same field carries different values in different materials. The quantity recorded on an acceptance record and the ordered quantity recorded on a purchase order concern the same subject, the same period, and the same field, so mutually exclusive values there form a contradiction. When one document is a purchase order and the other is a loss record, the fields differ, and the only honest result is a note that verification is pending. The same discipline applies to authorization states: a pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. At retrieval time, the vector model identifier is checked first, non-matching bases are excluded, and the actual observation is recorded. A cross-model retrieval is not run to compare missed recalls.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process inside hybrid retrieval, not a switch that can be turned on. The minimum relevance value applies only under semantic retrieval or when result re-ranking is used, and it acts as a filter threshold. It does not perform permission filtering or entity isolation. The citation limit is measured in tokens and controls how much content a single search may cite. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires image input and model capability. Re-ranking requires a re-rank model and must be enabled in settings. The node outputs cited content as an array whose length may be 0, and the downstream path still runs when nothing is retrieved.
A hypothetical paper exercise sample follows. A store fresh-produce supervisor submits a draft sentence for the conclusion paragraph of an exception review. Three source sentences are available for checking. The first comes from a purchase order: "On 3 June 2025, Store A ordered 200 boxes of peaches from Stall B, at the price of the day, settled after acceptance." The second comes from an acceptance record: "On 3 June 2025, Store A actually accepted 186 boxes of peaches, of which 14 boxes were downgraded for bruising." The third comes from a loss record: "On 4 June 2025, Store A recorded a loss of 6 boxes of peaches, with the cause noted as a storage temperature-control fault." The draft sentence reads: "Store A completed the purchase of 200 boxes of peaches on 3 June 2025, acceptance showed no issues, and loss was zero." The draft sentence and the acceptance record plus loss record carry mutually exclusive values in the same fields, so the draft is judged unsupported.
Configuration and implementation checks
The table below lists documented settings and implementation checks relevant to this feature and this review task, using the sample above. The numeric entries are starting points for this hypothetical exercise, not official defaults. The reason for the citation-limit starting point is that the sample materials include purchase orders and test reports of very different lengths, and the aim is to observe whether cited content stays complete. The comparison object is the content returned for the same question at different starting points.
| Configuration item | Hypothetical starting point for this exercise | How to check |
|---|---|---|
| Associated knowledge bases | Associate only the purchase, acceptance, and loss bases that use the same vector model | Check the vector model identifier and exclude bases on other models |
| Search mode | Run semantic retrieval and full-text retrieval separately first, then compare hybrid retrieval | Use the same question and observe the differences among the three paths |
| Citation limit | Set a starting token amount that covers the full text of one acceptance record | Compare against the token amount needed for the full purchase order text |
| Minimum relevance | Mark relevant and irrelevant passages in the sample first, then set a starting value | Observe whether the key acceptance-record sentence survives filtering |
| Question optimization | Enable for follow-up questions containing references, disable for explicit identifier questions | Compare the subject before and after optimization and check whether it was altered |
| Result re-ranking | Enable when candidates are numerous, and filter by the re-rank score | Observe the ordering difference between re-rank scores and semantic scores |
Three common errors
- Treating a retrieval hit as an authorization credential. The input is a retrieved purchase order; the comparison object is the approval status field. When that field is not stated, the only observable result is that the claim is unsupported, and no inference that the purchase was approved can be drawn.
- Treating the minimum relevance value as a permission or entity isolation mechanism. The input is a question spanning stores; the comparison object is the store identifier in the returned content. If materials from another store are returned, the observation is that this setting does not perform isolation, and access control must be implemented separately.
- Treating a pending review request as already executed. The input is a handover request marked pending review; the comparison objects are three distinct facts: signing, cooperation reached, and contract in effect. With only a pending marker, the observable result is that the authorization state is incomplete, and no executed conclusion follows.
How to verify the setup
- Verification one: Input the three source sentences and the draft sentence given above, then compare the quantity, acceptance status, and loss quantity in the draft against the values in the same fields in the source text. The condition is met when the draft is found to carry mutually exclusive values against the acceptance record and the loss record in the same fields and is judged unsupported.
- Verification two: Input a follow-up question containing a reference, then compare the subject before and after question optimization. The condition is met when the subject is not altered after optimization; if the subject is replaced, record it as a deviation introduced by question optimization.
- Verification three: Input a paper-level empty result array, then compare whether the downstream node still runs. The condition is met when the cited content has length 0 and the path continues to execute. This check does not claim that a search for any invented rule must return an empty array. All of the above are paper checks; no deployment or live test has been performed.