What the documents look like
Opening support for franchisees rests on four document families: training attendance sheets, opening checklists, SOP manuals, and assessment results. In cross-border programs these materials often exist in several language versions at once. A single store's attendance record, material-verification result, and corrective-action note may need to be compared across languages, and a retrieval task then requires that the connected model return the matching file segment by language for supervisors and trainers to review. The brief gives a per-store, monthly update rhythm; this page does not derive batch synchronization or real-time triggers from that rhythm.
The business difference from franchise filing management matters here. Opening support transfers operating capability; it is not legal disclosure. The typical parties are a chain training department, regional supervisors, and franchise store owners. The brief lists opening failure, inconsistent customer experience, and franchisee complaints as consequences of error, which is why the retrieval step must keep document state intact rather than smooth it into a finished result.
Constraints at the model connections step
AI Proxy is a required dependency. The App and Admin services must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail. Legacy OpenAI environment variables and single-model custom request addresses do not replace these two settings. The model ID is the value of the model field in the request body and is globally unique. Model mapping can map the model FastGPT requests to the actual upstream model, as in mapping gpt-4o-test to gpt-4o. The system needs at least one language model and one index model to work normally.
Retrieval behavior depends on the vector store. When the vector store is Milvus, full-text retrieval uses BM25 on the modeldata_v2 single table; other vector stores use MongoDB $text. Libraries built on different vector models cannot be associated together; the identifier must be checked first and non-matching libraries excluded. A relevance score does not carry permission or entity-isolation duties, and retrieval does not constitute authorization.
Hypothetical paper exercise sample. A chain training department submits a multilingual opening checklist. The source text reads: "Store number A-07; inspection date March 2026; inspection item: cashier counter materials complete; conclusion: pending review; reviewer has not signed." A draft sentence is offered for judgment: "The store's cashier counter materials were complete in March 2026 and have passed review." In the sample the reviewer has not signed, so the draft sentence converts a pending review into a passed review and is unsupported.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to this sample. Numeric entries are marked as hypothetical starting points for this exercise, with the comparison target stated.
| Configuration item | Hypothetical starting point for this exercise | Check note |
|---|---|---|
| AIPROXYAPIENDPOINT | Assume http://aiproxy:3000 | Root address with no /v1 appended; compare against the startup validation requirement |
| AIPROXYAPITOKEN | Assume a non-empty administrator token | Must match the server-side configuration; an empty value triggers startup failure |
| Model mapping | Assume gpt-4o-test maps to gpt-4o | Check that the FastGPT request model corresponds to the actual upstream model |
| Index model | Assume one index model is enabled | Compare against the requirement of at least one language model and one index model |
| Vector store type | Assume Milvus | Check whether full-text retrieval uses BM25 or MongoDB $text |
| Model probe | Assume a probe interval of 10 minutes | Compare against the documented integer range of 5 to 60 minutes |
Three common errors
- Writing
AIPROXY_API_ENDPOINTas a complete request address that includes/v1. Check: enter that address and compare it against the documented instruction to append no/v1; a startup validation failure or an abnormal request path means the condition is not met. - Writing the draft sentence for a pending-review checklist as a passed review. Check: enter the sample source text and the draft sentence, compare the reviewer signature state, and find the draft claiming a passed review while the source shows no signature; the condition is not met.
- Treating a relevance score as a permission filter. Check: enter the same query and compare returned records against the authorized scope; the appearance of unauthorized records means the condition is not met.
How to verify the setup
- Enter the sample source text and the draft sentence, compare the field values "pending review" and "passed review," and confirm the draft matches the source state. The current sample does not satisfy this.
- Enter the hypothetical
AIPROXY_API_ENDPOINTvalue, compare it against the documented root-address requirement with no/v1appended, and confirm the address format is correct and startup validation passes. - Enter a retrieval request across libraries built on different vector models, check the identifier first and exclude non-matching libraries; when identifiers differ, record the libraries as not associable and do not run a cross-model retrieval to compare missed recalls.
These checks describe expected observations on paper; no test, deployment, or approval has been performed. Where a document field is not provided, it remains unknown.