What the documents look like
Fruit-store procurement records sit in four document families. Purchase orders carry the item name, stall quotation and ordered quantity. Acceptance sheets carry the arrival grade and the weight deduction result. Pesticide-residue test reports are issued per batch. Loss records are registered by store and by date. The brief marks these updates as daily or per batch. Product-selection questions need to retrieve grade, origin, quotation and test status for one fruit category, and when the record subject, batch or date differ, retrieved results cannot be merged directly. Wholesale-market quotations and supplier quotes form the selection basis, while store sales data indicates replenishment direction. The category differs from convenience-store fresh-food ordering because it relies more on wholesale-market conditions and grade acceptance than on a headquarters standard meal.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together; the vector model identifier must be checked first, non-matching bases excluded, and the actual observation recorded. Cross-model retrieval is not run to compare missed recalls. Retrieval treats the knowledge base as the smallest search unit; a collection only classifies data and does not affect search results. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process, not a standalone switch. Semantic retrieval and the minimum relevance score do not act as permission or entity isolation; retrieved results do not constitute authorization. The citation limit is measured in tokens, and the cited-content output can be an array of length 0, meaning the downstream path still runs when nothing is found. Image retrieval requires image input and matching model capability; reranking requires a model and must be enabled by configuration. The paths drawn in the overall flow diagram do not mean every request executes all of them. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval.
Configuration and implementation checks
The table lists documented settings or implementation checks for this page. The values are hypothetical starting points for this trial, chosen for comparison against observed results, and are not official defaults.
| Configuration item | Hypothetical trial starting point and reason | Check method |
|---|---|---|
| Associated knowledge bases | Associate one knowledge base containing purchase orders, acceptance sheets, test reports and loss records; first confirm the vector model identifier is consistent | Compare base identifiers and vector model identifiers; exclude inconsistent ones |
| Search mode | Start with hybrid retrieval to cover both item codes and natural-language questions | Compare retrieved content using a code-type question and a description-type question |
| Citation limit | Set the starting point to cover the full fields of one acceptance sheet, so fields are not truncated | Compare citation content completeness using records of different lengths |
| Minimum relevance | Label relevant and irrelevant samples first, then set a starting point and compare filtered results | Observe whether filtered content matches the labelled samples |
| Question optimization | Enable for multi-turn follow-ups and short questions; a single clear question may keep it off | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when candidates are many and the question is clear; check the rerank score | Compare ordering before and after reranking and record the score range |
Three common errors
- Treating semantic retrieval or the minimum relevance score as a permission filter. In a hypothetical drill sample, a buyer is authorized to view only this store's acceptance sheets, yet retrieval returns records from other stores. The error is using a similarity threshold as access control. The check input is the retrieval account and the target base scope; compare the store field of returned records against the authorized scope, and observe whether out-of-scope records appear.
- Associating bases that use different vector models. In a hypothetical drill sample, two bases use different vector models and the operator selects both directly. The check input is the vector model identifier of each base; compare whether the identifiers match, and observe whether the base is excluded or recorded as not associable.
- Treating a purchase request awaiting review as approved and executed. In a hypothetical drill sample, one purchase order is in a pending-review state, and the draft sentence reads "this batch of fruit has completed procurement." The check input is the purchase order status field; compare the status value against the draft sentence wording, and observe whether the sentence exceeds the recorded status.
How to verify the setup
Hypothetical drill sample. The acceptance sheet reads: "On 12 March 2025, Store A accepted one batch of mandarin oranges supplied by stall B, arrival grade two, net weight after deduction 300 kilograms, accepted by Zhang." The pesticide-residue test report reads: "On 12 March 2025, the mandarin orange sample of batch C tested qualified, report number D." The draft sentence to be judged is: "Store A has completed procurement of this batch of mandarin oranges and they may be put on shelves for sale."
- Citation completeness check. Input the acceptance sheet text above and the draft sentence. Compare whether the draft sentence uses only the store, date, item name, grade and net weight stated in the original, and observe whether the draft sentence adds a "may be put on shelves for sale" conclusion not stated in the original.
- Status field check. Input the purchase order status field above and the draft sentence. Compare the status value against the draft sentence wording, and observe whether the sentence describes a pending-review state as completed procurement.
- Empty-result downstream check. Input the paper value [] and compare whether the downstream design continues execution according to a cited-content length of 0. Observe whether the downstream treats the empty result as no authorization or as a retrieval failure. An empty result only means no citation was returned this time; it does not mean a fictional rule retrieval must return an empty array.