What the documents look like
Supermarket supplier entry settlement runs on purchase-and-sale or consignment payables, and a supplier does not necessarily operate a staffed counter on site. The working material sits in four document families: purchase contracts, delivery acceptance notes, rebate agreements and reconciliation statements. Product information and selection question answering must answer what settlement conditions apply to a given item under a given contract, which accepted-quantity basis applies, and which rebate deduction basis is supported. The answers draw on contract clauses, acceptance records and agreement annexes, and the reference material moves on a weekly or monthly cycle. Answers need to stay consistent with the reconciliation statement for the same period.
Hypothetical drill sample. A supplier and a supermarket sign a purchase-and-sale contract, contract number HT-2026-031, signed on 2026-03-01, agreeing that settlement terms for one item run 30 days after acceptance, and that the rebate is calculated at 2% of quarterly sales. Delivery acceptance note YS-2026-0412 records 800 units accepted on 2026-04-12, of which 20 units were returned because the outer packaging was damaged. Rebate agreement FL-2026-007 states that the quarterly rebate uses the actually accepted qualified quantity as its base. One draft sentence awaiting judgment reads: "The settlement quantity for this item in April 2026 is 800 units, and the rebate base is 800 units."
Constraints at the model connections step
Model connections depend on AI Proxy. From V4.17.0, both the App and any Admin service that reuses App/Service configuration must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN; a missing item causes startup validation to fail. AIPROXYAPIENDPOINT holds the service root address and does not append /v1. AIPROXYAPITOKEN is an administrator token and an application API key does not substitute for it. The system needs at least one language model and one index model to work normally.
Model capability is checked against the input the application requires. Multimodal input, tool calling and thinking configuration are each verified separately, and parameters are set in the application or node configuration. A same-named field on a compatible chat interface does not overwrite orchestration settings.
For the drill sample, the model connection must let a question about item settlement conditions, acceptance quantity basis and rebate deduction basis be answered from the contract, acceptance note and rebate agreement. Retrieval is not authorization. A relevance score is not a category, entity or permission filter, and a similarity setting does not exclude unrelated businesses or unauthorized records. Same-model retrieval association keeps its documented condition: libraries built on different vector models cannot be jointly associated, so the identifiers are checked first and non-matching libraries are excluded before any observation is recorded.
Configuration and implementation checks
The table below covers only settings and implementation checks tied to this page's documents and task. Each starting value is a hypothetical drill starting point, not an official default.
| Configuration item | Hypothetical drill starting point | Reason and comparison object |
|---|---|---|
| AIPROXYAPIENDPOINT | http://aiproxy:3000 | Documented example form for the service root address; compared against a value carrying a /v1 suffix |
| AIPROXYAPITOKEN | A non-empty administrator token entered for the drill | Compared against an empty value and against an application API key record, to confirm startup validation passes |
| Language model | One tested language model enabled | Compared against a disabled state, to confirm a question can be issued |
| Index model | One tested index model enabled | Compared against a disabled state, to confirm retrieval is available |
| Model test | A model test run against the channel | Compared against a failed test result, to confirm the channel is valid |
| Model mapping | Mapping entries filled per channel | Compared against the requested model ID, to confirm the upstream receives the actual model |
The language model and index model entries are the minimum the documented material states for normal use. Enabling a model, running a channel test and reading a test result are operator-visible actions in the administration pages. Business verification, approval, database writes and external synchronization are not part of model connection configuration and require their own downstream implementation. A pending request is not approval, and an offered option is not execution.
Three common errors
- Writing AIPROXYAPIENDPOINT as a full request address carrying
/v1. Check: read the configured value, compare it against the documented service root address form, and a/v1suffix makes it non-conforming. Input is the configuration value; comparison is the documented root address form; observable result is the presence of the suffix. - Using an application API key as AIPROXYAPITOKEN. Check: read the token source identifier, compare it against the application API key record, and a source mismatch makes it non-conforming. Input is the token source label; comparison is the application API key record; observable result is the source mismatch.
- Treating the 800 units in the draft sentence as both the settlement quantity and the rebate base. Check: read delivery acceptance note YS-2026-0412, compare it against the draft sentence quantity, and a draft that does not deduct the 20 returned units does not match the accepted qualified quantity basis. Input is the acceptance note and the draft sentence; comparison is the stated quantity against the qualified quantity; observable result is the undeducted figure.
How to verify the setup
- Input the draft sentence from the hypothetical sample together with the original text of delivery acceptance note YS-2026-0412. Compare the 800 units in the draft sentence against the qualified quantity of 780 units in the acceptance note. Where the draft does not reflect the returned-unit deduction, the draft is judged as needing correction. This is an expected judgment on paper; no test has been run.
- Input the sentence in rebate agreement FL-2026-007 stating that the quarterly rebate uses the actually accepted qualified quantity as its base. Compare the 800-unit rebate base in the draft sentence against that clause's base basis. Where the two differ, the draft is judged as not supported. The agreement text is the comparison source.
- Input the settlement term clause of contract HT-2026-031. Compare whether the draft sentence states the settlement time point. Where the draft does not mention the settlement term, record it as an evidence gap and keep the item pending verification. The contract clause is present in the drill material; the draft's silence is the observable condition.
For each check the input, the comparison and the observable result are stated. Where the material does not carry a field, the field remains unknown. A pending verification item is not a negative finding and is not a conflict; it stays as an evidence gap until the missing clause or record is supplied.