What the documents look like
Store count and loss-control work centers on book-to-actual variance and shrinkage. It does not settle supplier payables. The working material comes from POS inventory, WMS, camera systems, and the SB/T count specification. Day-to-day documents include count sheets, variance analyses, write-off approvals, and camera review records. The typical roles are the loss-prevention department, store inventory keeper, and finance auditor. Data refresh runs monthly or quarterly; this page's feature does not depend on that cadence.
Product-information and selection questions ask how a given SKU's book quantity, actual quantity, and variance-reason field are valued in a count sheet, and which documents a selection configuration relies on. Because the source material updates slowly, one count sheet may be cited repeatedly over several months, so a configuration must be able to distinguish versions.
Constraints at the model connections step
Model connections depend on AI Proxy. Both App and Admin must set AIPROXY_API_ENDPOINT (the service root address, without an appended /v1) and a non-empty AIPROXY_API_TOKEN; missing either one fails the startup check. Legacy OpenAI environment variables or a single-model custom request address cannot substitute for these two. Model capabilities are checked against the input the application needs: multimodal, tool calling, and thinking configuration are verified separately. Application parameters are set in the application or node configuration; a same-named field on a compatible chat interface is not assumed to override the orchestration.
Hypothetical paper sample: a store count sheet dated 2026-03-31 states: "Count date 2026-03-31, SKU-A book quantity 120 units, physical count 108 units, variance -12 units, variance-reason field blank, write-off approval number not filled." A separate variance-analysis draft reads: "SKU-A variance 12 units, confirmed as normal shrinkage, no write-off approval required." The question is whether that draft sentence is supported by the count sheet. The sheet states no variance reason and no write-off approval status. The draft's "confirmed as normal shrinkage" and the sheet's blank variance-reason field are not mutually exclusive values in the same field, so the claim is unsupported; "no write-off approval required" has no corresponding field in the source, so it is pending verification. This sample is for paper checking only and represents no real store or actual test.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | http://aiproxy:3000 | Official example root address, no /v1 appended. Compare character by character against the documented example to confirm no path suffix |
| AIPROXYAPITOKEN | A non-empty string | Official requirement is non-empty and matching the server side. Compare against the AI Proxy administrator token for equality |
| Language model | One enabled model | Official requirement is at least one language model. Check enabled status on the system model page |
| Index model | One enabled model | Official requirement is at least one index model. Check enabled status on the system model page |
| Model probe interval | 30 minutes | Official range is an integer from 5 to 60. Compare against the range boundaries to confirm the value falls inside |
| Model capability fields | Checked against the input the question needs | Official guidance is to check against the application's required input. Inspect each field such as vision and toolChoice for a match |
Three common errors
- Entering a full model request address (containing
/v1or/chat/completions) asAIPROXY_API_ENDPOINT. The official text requires the service root address. The check takes the variable value as input, compares it for a path suffix, and treats any suffix as not satisfying the condition. - Using an application API key in place of
AIPROXY_API_TOKEN. The official text requires an administrator token that is not empty. The check takes both values as input, compares them for equality, and treats a mismatch as not satisfying the condition. - Reading "variance-reason field blank" in the hypothetical sample as "confirmed as normal shrinkage." The check takes the count sheet text and the draft sentence as input, compares the variance-reason field value, and finds the source blank while the draft asserts "confirmed," so the assertion is unsupported.
How to verify the setup
- Take the hypothetical sample's count sheet text and draft sentence as input, and compare the draft's "confirmed as normal shrinkage" against the source's variance-reason field. The source field is blank, so the draft sentence is unsupported; if the draft instead read "variance reason pending verification," it would match the source. The expected observation is that the judgment changes with the source value.
- Take the hypothetical sample's write-off request "SKU-A requires no write-off approval" as input, and compare it against the source's write-off approval number field. The source leaves it unfilled, so the request remains pending verification and does not support a conclusion that approval was completed or that approval is absent. The expected observation is that the evidence gap is preserved.
- Take two candidate values for
AIPROXY_API_ENDPOINT(one with/v1, one without) as input, and compare them against the official requirement. The value without a path suffix satisfies the condition; the value with the suffix does not. The expected observation is that the configuration check result differs.
The 48-hour stability display, the four-attempt abnormal determination, and the edge-triggered Webhook are mechanisms described in the official documents; this page does not claim they were run. Retrieval and relevance scores carry no permission or entity isolation. Libraries built on different vector models cannot be jointly associated; check identifiers and exclude non-matching entries.