What the documents look like
Franchisee training and opening-coaching material comes from the training system, the supervisor app, store photographs and the headquarters SOP library. The working documents are training sign-in sheets, opening checklists, SOP manuals and assessment scores, refreshed per store per month. Product information and selection Q&A handle retrieval and citation of this material after model connections are in place. The material carries store, batch and date scope. Retrieval returns text fragments only and does not change the authorization state of any record.
Two features of this category shape the model-connection step. First, opening coaching transfers operating capability rather than filing a legal disclosure, so the evidence that matters is training and inspection records rather than registration documents. Second, the same store appears in several documents with different dates and different fields, so a statement about one store can be supported by one record and unsupported by another. A sign-in sheet and an opening checklist for the same store on different dates answer different questions.
Constraints at the model connections step
From V4.17.0, AI Proxy is a required dependency. The app and any Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; a missing item causes startup validation to fail. Legacy OpenAI environment variables or a single-model custom request address do not replace these two settings. The system needs at least one language model and one index model to work normally. Only one speech-recognition model takes effect, so configuring one is sufficient.
Hypothetical drill sample. A chain brand training department intends to check one opening-coaching item. The first source text reads: "March 2026 batch, East China franchisee opening checklist, inspection date 10 March 2026, inspector regional supervisor Li, store number HD-0310, items cover cashier operation and inventory counting, inspection conclusion pending rectification." The second source text reads: "March 2026 batch, East China franchisee training sign-in sheet, sign-in date 8 March 2026, attendee HD-0310 store owner Wang, training content SOP manual chapter three cashier process, sign-in status signed." The draft sentence to be judged reads: "Store HD-0310 has completed opening rectification and meets opening conditions." This draft must be checked against the checklist conclusion and the sign-in status.
The checklist records a conclusion of pending rectification. It does not record that rectification was completed or that opening conditions were met. The sign-in sheet records attendance at training on 8 March 2026. It does not record opening conditions. Retrieval returning these fragments means the text was recalled; it does not mean rectification finished, opening was approved, or any business action was executed. A pending review does not establish that operating permission is absent, and a pending rectification does not establish that rectification has not started.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to this page's function and primary task. The starting values are hypothetical drill starting points, not official defaults.
| Configuration item | Hypothetical drill starting point | Check note |
|---|---|---|
AIPROXY_API_ENDPOINT | Set to http://aiproxy:3000 | Documented as the AI Proxy service root address, without /v1 appended. This starting point checks that app and Admin agree. |
AIPROXY_API_TOKEN | Set to a non-empty administrator token matching the server | Documented as non-empty and matching the server; an application API key does not substitute. |
| Language model and index model | One of each enabled | Documented as the minimum for normal operation. This starting point checks that both categories are enabled. |
| Model capability | Checked against the input the application needs | Multimodal input, tool choice and thinking configuration are verified separately. The sample is text Q&A and does not involve image input. |
| Application parameters | Set in the application or node configuration | Parameters are set in the application or node configuration; a same-named field on a compatible chat interface is not assumed to override orchestration. |
| Model status probe | Probe interval set to an integer between 5 and 60 minutes | The documented range is an integer in that interval. This starting point checks whether enabled models participate in probing. |
Three common errors
- Treating the sign-in sheet status "signed" as evidence that opening rectification is complete. The sign-in sheet records training attendance on 8 March 2026. The checklist records an inspection on 10 March 2026 with a conclusion of pending rectification. The subject is the same, the periods and the fields differ, and the sign-in status does not support "rectification completed".
- Reading the checklist conclusion "pending rectification" directly as "meets opening conditions". The checklist source records only a conclusion of pending rectification. It states neither completed rectification nor satisfied opening conditions. Both "rectification completed" and "meets opening conditions" in the draft sentence remain unsupported.
- Treating a retrieval hit as authorization or execution. A returned fragment means the text was recalled. It does not mean rectification finished, opening was approved, or any business action was carried out. A pending review does not establish that operating permission is absent, and a pending rectification does not establish that rectification has not begun.
How to verify the setup
- Input the checklist source text and the sign-in sheet source text above. Compare "rectification completed" in the draft sentence "Store HD-0310 has completed opening rectification and meets opening conditions" against the checklist conclusion field. The checklist conclusion value is "pending rectification" for the same subject and the same field. The observed result is that this part of the draft is unsupported by the checklist, and the evidence gap is recorded.
- Input the sign-in sheet source text. Compare "meets opening conditions" in the draft sentence against the recorded sign-in status. The sign-in sheet records training attendance only and states no opening condition. The observed result is that the sign-in sheet does not constitute evidence of opening conditions, and this part remains to be verified.
- Input both source texts and the draft sentence. Compare the material date scope against the period the draft sentence concerns. The checklist date is 10 March 2026 and the sign-in sheet date is 8 March 2026, while the draft sentence states no specific period. The observed result is that, until a rectification-completion record or an opening-approval document is supplied, the draft sentence stays unsupported. Where retrieval returns no result, a plain paper input
[]can be used to check downstream design; this does not assert that a fictional rule always returns an empty array. A relevance score is not a category, entity or permission filter, and retrieval is not authorization.