What the documents look like
Fresh-food concession counters in hypermarkets are often staffed and operated by suppliers on site, with revenue shared by sales. Product information, daily sales reports, loss sheets and food incoming inspection records come from the store's fresh-food department, the concession supplier and store stocktaking. These materials do not naturally align on subject, date or quantity scope. The concession contract states the revenue-share and settlement arrangement. The daily sales report records weighed sales. The loss sheet records reported write-off quantity. The food incoming inspection record records batch documentation. Product information and selection questions require aligning fields from these documents to the same subject, the same period and the same field before workflow routing decides when to collect information from the user, when to let the user make a choice, and when to pass results to later nodes.
Constraints at the workflow routing step
When the form input node triggers, the conversation enters an interaction state, collects required information, and passes that information to later nodes after submission. When the user selection node triggers, the conversation also enters an interaction state, and after the user makes a choice the corresponding "yes" branch executes. Both nodes record workflow state and wait for the user to complete interaction before continuing execution. The vector model identifier of a retrieval library must match the query side; libraries using different vector models cannot be associated together. A relevance score indicates similarity only and does not perform permission or entity isolation. Order writing, document backfill and external state synchronization require a separate business interface or later process; the nodes themselves do not complete these business actions.
Hypothetical drill sample: a concession contract between a hypermarket fresh-food department and a concession supplier states "the settlement period is the calendar month, and loss is shared equally between both parties." The same month's daily sales report records "weighed sales amount 12,000 yuan." The loss sheet records "reported write-off 300 yuan." The food incoming inspection record states "March 5, 2025 incoming batch A, quantity 200 kilograms." A proposed sentence reads "the supplier's loss share for this month is 150 yuan." This sentence multiplies the loss sheet amount by the contract share ratio, but the loss sheet does not state whether the 300 yuan falls within the contractually agreed share scope, nor whether non-shareable portions have already been deducted.
Configuration and implementation checks
| Configuration item | Relevant value or check for this page | Explanation |
|---|---|---|
| Form input node | Enters interaction state on trigger, collects required information | Passes information to later nodes after submission |
| User selection node | Executes the corresponding "yes" branch after the user makes a choice | Check branch connections one by one |
| Interaction state check | Recording state is behavior of the interaction node | No separate global state-recording switch is set |
| Output handoff check | Passes collected information to later retrieval, judgment or interface nodes | Compare output against input item by item |
| External operation division | Order writing, document backfill and external state synchronization | Require a separate business interface or later process |
| Retrieval library model identifier | Matches the query-side vector model identifier | Libraries using different models cannot be associated together |
Three common errors
- Treating the loss sheet amount collected by the form input node as the contractual share base. The loss sheet records only reported write-off quantity or amount and does not state whether that amount falls within the contractually agreed share scope. To check, input the loss sheet text and the contract share clause, compare whether both point to the same cost scope, and if the loss sheet does not state the share scope, record it as unsupported.
- Treating the "yes" branch of the user selection node as approved execution. User selection indicates only that the user made a choice; signing, reaching cooperation and contract effectiveness are three distinct facts. To check, input the user selection record and the proposed sentence, compare whether the proposed sentence writes the selection result as a completed action, and if the proposed sentence contains "approved" or "effective," record it as pending verification.
- Treating a retrieval relevance score as permission or entity filtering. A relevance score indicates similarity only and does not exclude unrelated business or unauthorized records. To check, input the retrieval return result and the proposed sentence, compare whether the proposed sentence claims the score has filtered unauthorized records, and if such a claim appears, record it as beyond the documented description.
How to verify the setup
- For the proposed sentence "the supplier's loss share for this month is 150 yuan" in the hypothetical drill sample, input the loss sheet text, the contract share clause and that proposed sentence, and compare whether the proposed sentence matches the cost scope stated in the loss sheet. If the loss sheet does not state that the 300 yuan falls within the share scope, the proposed sentence is unsupported and the evidence gap must be recorded.
- For the user selection node, input a hypothetical interaction record explicitly marked as hypothetical, in which the user selects "yes," compare whether the later node executes along the "yes" branch connection, and check whether the proposed sentence writes that selection as approved or effective. If the proposed sentence contains executed wording, record it as pending verification.
- For the retrieval library model identifier, input the query-side model identifier and the retrieval library model identifier, and compare whether the two match. If the identifiers do not match, exclude that library and record the actual observation; do not perform cross-model retrieval to compare missed recall. Empty-result handoff can be checked against the downstream design using the paper input
[], without claiming that a fictional rule retrieval must return[].