What the documents look like
Supermarket supplier onboarding and settlement centers on purchase-sale or consignment payables, and in most cases there is no resident counter operation. Records arrive from ERP procurement, supplier portals, POS sales and finance systems. The everyday documents are purchase contracts, delivery acceptance notes, rebate agreements and reconciliation statements. In cross-border and multilingual settings, one acceptance note may carry field descriptions in Chinese, English or a third language at the same time. Rebate clauses in a purchase contract and the deduction basis on a reconciliation statement must be compared across documents. Supplier business staff, supermarket buyers and finance settlement staff each hold a fragment of the information, and the material scope any one party can retrieve may differ. These characteristics shape what model connection and configuration must achieve: make multilingual material locatable during retrieval, and treat a retrieval hit as a location result rather than an authorization credential.
Two or three industry facts matter most here. The first is that settlement documents are split across procurement, portal, POS and finance sources, so a single question may need evidence assembled from more than one system. The second is that multilingual fields describe the same commercial fact in different languages, which makes cross-document comparison a normal step rather than an exception. The third is that three separate parties hold partial views, so a retrieved fragment reflects only the material scope available to that retrieval, not the full authorization picture. Weekly or monthly update frequency does not change the model connection step described on this page.
Constraints at the model connections step
From V4.17.0, AI Proxy is a required dependency. The App service and any Admin service that reuses App/Service configuration must both set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN; missing either one causes startup validation to fail. Legacy OpenAI environment variables or a single-model custom request address cannot replace these two settings. AIPROXYAPIENDPOINT takes the service root address and must not carry a trailing /v1. The system needs at least one language model and one index model to work normally. Only one speech recognition model takes effect. The model ID is the value of the model field in the request body and is globally unique; the model name is customizable for display. One model can be configured across different channels to achieve load balancing. Channel priority ranges from 1 to 100, and a larger value is more likely to be selected first. Model probing covers only currently enabled system models; after a first failure the system retries automatically at most 3 times, and only when all 4 attempts fail is the model judged abnormal. Retrieval is the process of locating material and does not carry permission or entity isolation. Knowledge bases that use different vector models cannot be associated together.
Configuration and implementation checks
The following is a hypothetical paper exercise used to explain how configuration is checked. It does not represent any real record and does not indicate that testing has been completed.
Hypothetical paper exercise: in a supermarket supplier settlement setting, a Chinese purchase contract states "rebates are settled quarterly, based on accepted quantity." An English delivery acceptance note states "accepted quantity: 1200 cartons." A reconciliation statement states "this quarter's rebate deduction is 3%." A draft sentence awaits judgment: "This supplier's quarterly rebate has been settled at the accepted quantity and a 3% deduction." The material available for checking covers only the contract clause, the accepted quantity and the deduction rate; it does not state that settlement has been completed.
| Check item | Hypothetical starting value for this exercise | Reason for the choice and comparison object |
|---|---|---|
| AIPROXYAPIENDPOINT | Assume http://aiproxy:3000 | Matches the documented example; compare whether /v1 was added by mistake |
| AIPROXYAPITOKEN | Assume a non-empty administrator token | Matches the documented requirement; compare whether it is empty or an application API key |
| Language model and index model | Assume one of each is enabled | Compare against the documented minimum requirement that both types exist |
| Channel priority | Assume 50 | Compare against the documented 1 to 100 range and confirm it falls inside |
| Model probe interval | Assume 30 minutes | Compare against the documented 5 to 60 minute integer interval |
| Model ID and model name | Assume model ID gpt-4o-test with a display alias as the name | Compare whether the two fields are being treated as one |
Three common errors
- Filling AIPROXYAPIENDPOINT with a full model request address and appending /v1. The input is the configuration text; the comparison object is the documented "service root address, do not append /v1." If the address ends in /v1 or /chat/completions, the condition is not met.
- Placing an application API key into AIPROXYAPITOKEN. The input is the content of the token field; the comparison object is the documented "administrator token, must match the server configuration and cannot be empty." If an application-side credential is entered, the condition is not met.
- Treating a retrieval hit as evidence of authorization or of completed settlement. The input is the material fragment returned by retrieval; the comparison object is whether the source text states an authorization or a completed settlement fact. If the source text only gives the rebate rate and accepted quantity without stating completion, the phrase "has been settled" in the draft sentence is unsupported.
How to verify the setup
The three checks below all use the hypothetical paper exercise above. The expected results are paper-based conclusions and were not actually executed.
- Check AIPROXYAPIENDPOINT and AIPROXYAPITOKEN. The input is the hypothetical configuration text. Compare whether the address is a root address and whether the token is non-empty, checking the root address format, token non-emptiness and its consistency with the server. This checks only two configuration conditions; overall startup and invocation results remain untested. When a required item is missing, record the startup validation gap as documented.
- Check the draft sentence "This supplier's quarterly rebate has been settled at the accepted quantity and a 3% deduction." The input is the draft sentence and the three source texts. Compare the same subject, the same period and the same field. The material states the rebate basis and the deduction rate, and does not state that settlement has been completed. The phrase "has been settled" is therefore unsupported; record the evidence gap and do not judge it as completed.
- Check model probe status. The input is the hypothetical probe configuration and probe results. Compare whether the probe scope covers only enabled models and whether a single round's attempt count falls between 1 and 4. If a model fails all 4 attempts, its status is abnormal; if it succeeds with latency not exceeding 30s, its status is normal. Disabled models do not participate in probing and produce no history records.