What the documents look like
Brand after-sales repair work turns on warranty responsibility and physical repair locations. It differs from platform refund handling, which centers on reversing a transaction. The material base combines the after-sales system, store orders, repair-location feedback, and the national three-guarantees catalogue. Daily documents cluster around warranty cards, repair work orders, inspection reports, and replacement approvals. Product information and selection questions must state warranty scope, spare-part matching, and approval status clearly, so model connection and configuration checks land on the fields inside these documents. Records update per order and daily; that cadence affects when freshness is checked and does not change the connection method on this page.
Constraints at the model connections step
Model access depends on AI Proxy. Both the App and Admin services must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; a missing item fails startup validation, and legacy OpenAI variables or a single-model custom request address cannot substitute for these two. The system needs at least one language model and one index model. Only one speech-recognition model takes effect, so configuring one is sufficient. A model ID is the value of the model field in the request body and is globally unique; the model name is a display label that can be customized. Channel tests and page tests send real requests and serve as an initial check that a model works. Model status probing covers only currently enabled system models, retries up to three times after a first failure, and marks a model abnormal only when all four attempts fail. Retrieval and similarity scores carry no permission or entity-isolation duty, and the same-vector-model requirement for a single knowledge base does not extend to external systems.
Hypothetical drill sample. Warranty card text for comparison: "Warranty for this unit runs 12 months from 1 March 2025. Mainboard failure falls within warranty scope. A purchase invoice must be provided." Repair work order text: "Accepted 8 January 2026. Fault described as mainboard will not power on. Customer did not provide an invoice. Work order status is pending inspection." Draft handover sentence: "This mainboard failure is within the warranty period, so a free replacement can be arranged directly." Request to decide: submit the draft sentence as the reply content of a product-information question for human review.
Configuration and implementation checks
The table lists only documented settings or implementation checks tied to this page's main task. Values are hypothetical starting points for this exercise, with the reason and the comparison object stated.
| Setting or check | Hypothetical starting point | Reason and comparison object |
|---|---|---|
| AIPROXYAPIENDPOINT | Enter the AI Proxy service root address without appending /v1 | Compare against the documented instruction not to append /v1 |
| AIPROXYAPITOKEN | Enter a non-empty administrator token matching the server | Distinguish from an application API key; compare whether it is empty |
| Language model and index model | Enable one of each | Compare against the requirement of at least one language model and one index model |
| Model ID and model name | Model ID is globally unique; model name is customizable | Compare field by field against the channel model entry |
| Model status probe interval | Hypothetical starting point of 30 minutes | Documentation supports integers from 5 to 60 minutes; compare against both bounds |
| Retrieval scope check | Record the returns actually observed and an empty result | Compare against the downstream design when the input is [] |
Three common errors
- Mixing up
AIPROXY_API_TOKENwith an application API key. Input is the environment variable configuration item; the comparison object is the documented administrator token; the observable result is whether startup validation passes. - Treating the model name as the request identifier. Input is the display name cited in a draft sentence; the comparison object is the globally unique model ID; the observable result is the
modelfield in the request body during a channel test. - Treating a similarity score as a permission or entity filter. Input is one retrieval request; the comparison object is the documented retrieval duty; the observable result is whether returned records were checked against a separate authorization scope.
How to verify the setup
- For the hypothetical sample's draft handover sentence, input the warranty card text and the repair work order text, then compare three fields: warranty start and duration, whether an invoice was provided, and work order status. The condition is met when the draft sentence's conclusion can be directly supported by the source text. In this sample the invoice is missing and the work order is pending inspection, so the draft sentence is not supported and the evidence gap should be recorded.
- For model connection, input the
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENconfiguration values, then compare whether either is empty and whether/v1was appended. The condition is met when both are non-empty and the root address carries no/v1; when unmet, record it as a configuration gap. - For the retrieval check, input one query and a clearly specified empty result
[], then compare the returned records against the downstream design. The condition is met by recording the returns actually observed and separately checking the authorization scope; retrieval itself does not constitute authorization and does not guarantee a first-hit match or full recall.