What the documents look like
A brand concession inside a shopping mall accepts the mall's unified cashiering and settles on a concession-fee basis. Sales receipts originate from the mall POS, while the brand ERP and the finance system keep a separate set of records. Daily-frequency material comes from the mall POS and sales receipts; monthly-frequency material comes from the monthly settlement statement. The joint-operation concession contract and the mall event letter are clause-type materials: the contract sets the concession rate and settlement cycle, the event letter sets the resources and cost-sharing terms during the event period. Product data and selection questions require field-level reconciliation across these materials: whether the same concession, the same settlement period and the same concession-rate field carry consistent values.
This page addresses one task only: product data and selection questions at the model connection and configuration step. It does not treat retrieval results or form submissions as authorization to execute a settlement, an event charge or a deduction.
Constraints at the model connections step
Model access depends on AI Proxy. The App and any Admin service that reuses App/Service configuration must both set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; a missing item causes startup validation to fail. Old OpenAI environment variables or a single-model custom request address do not substitute for these two settings. The system requires at least one language model and one index model to operate normally. Model configuration is performed on the administrator model provider page; channel testing and model testing are separate actions after configuration.
The retrieval library and vector model identifier must be checked first. Libraries built on different vector models cannot be associated together; the identifier should be verified and non-matching libraries excluded, without running cross-model retrieval to compare missed recalls. A relevance score does not carry permission or entity isolation duties, and retrieval results do not constitute authorization. An empty result can be checked against the explicit paper input [] for downstream design; it cannot be claimed that a fictional rule retrieval necessarily returns [].
Hypothetical drill sample: the September 2026 monthly settlement statement for a brand concession states in its source text: "Concession No. A-021, settlement period 2026-09-01 to 2026-09-30, joint-operation concession rate 18%, event cost sharing 2,400 yuan." The mall event letter for the same month states in its source text: "Event period 2026-09-15 to 2026-09-17, event cost sharing cap 2,000 yuan." The draft sentence to be assessed is: "Concession A-021 September event cost sharing 2,400 yuan, consistent with the event letter cap."
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point and reason | What it is compared against |
|---|---|---|
AIPROXY_API_ENDPOINT | AI Proxy service root address, without appending /v1 | The actual deployed service root address |
AIPROXY_API_TOKEN | A non-empty administrator token | The AI Proxy server-side configuration |
| Language model | At least one enabled | The system minimum requirement |
| Index model | At least one enabled | The system minimum requirement |
| Vector library identifier | Record the vector model identifier used by this library | The identifier of the library to be associated; non-matching libraries are excluded |
| Model test | Execute item by item for each configured channel | The channel test result |
Three common errors
- Treating a retrieval hit as an authorization basis. A matched settlement statement fragment only shows the text exists; it does not show that the concession has been approved to execute a particular cost-sharing amount.
- Associating libraries across vector models directly. When two libraries use different vector models, verify the identifier and exclude non-matching libraries first; do not run cross-model retrieval to compare missed recalls.
- Writing a pending-review state as completed. If the event letter only states "pending mall confirmation," the cost-sharing amount is in a to-be-verified state and does not constitute approval.
How to verify the setup
- Input the settlement statement source text "joint-operation concession rate 18%" and the event letter source text "event cost sharing cap 2,000 yuan" from the hypothetical sample, and compare the draft sentence "cost sharing 2,400 yuan, consistent with the event letter cap." The compared fields are the event cost-sharing amount and the cap. Observed result: 2,400 yuan exceeds the 2,000 yuan cap, so the "consistent" wording in the draft sentence does not hold and should be recorded as to-be-verified.
- Input the event letter source text "event period 2026-09-15 to 2026-09-17" from the hypothetical sample, and compare the draft sentence's reference to "September event cost sharing." The compared fields are the event period and the settlement period. Observed result: the event letter covers only September 15 to 17, which differs from a full-month settlement period; the draft sentence equates event cost sharing with full-month sharing without support.
- Input the paper empty result
[]and compare the downstream question-answering output design when there is no retrieval hit. Observed result: downstream should present a no-hit state and not generate fictional clauses; if the output contains a concession rate or cost-sharing figure not stated in the materials, record it as a failure branch.
These checks describe expected observations for the named primary task. They do not report that any test, deployment, approval, refund or other business action actually occurred. Where a document field is not provided, the value remains unknown.