What the documents look like
Complaint and regulatory reply work draws on the 12315 platform, platform complaint centres, customer service chat and order systems. The material is addressed to a platform or a regulator, so it requires formal evidence and remediation wording rather than informal service talk. The documents handled day to day are complaint tickets, regulatory enquiry letters, order evidence and remediation reports. In the product information and selection question-answering task, these documents supply the product facts and the stated basis for a selection decision, and the model connection step has to let a language model and an index model read and answer from that content.
Model access depends on the AI Proxy service. The App and Admin services must both 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 system needs at least one language model and one index model before it can be used normally. The index model builds indexes over text chunks for related-text retrieval, and the language model produces the answer. Retrieval is a recall process and does not perform authorisation checks, so a similarity setting does not exclude unrelated businesses or unauthorised records. Libraries built on different vector models cannot be associated together; identifiers must be checked first and non-matching libraries excluded.
Constraints at the model connections step
The model connection decides whether question answering can cite product facts from complaint tickets and regulatory enquiry letters. The following is a hypothetical paper exercise, used only to show a checking method. It does not indicate that any test or deployment took place.
Material text one: a complaint ticket records that a consumer bought a particular model of rice cooker on 5 March 2025, order number A-20250305-017, and that the complaint reason is "the product page states a capacity of 5 litres, while the actual measurement is 4.2 litres".
Material text two: a regulatory enquiry letter requests that the complained party submit, within seven working days of receiving the letter, the purchase vouchers for that model and the basis for the page advertising claims.
Draft sentence to be assessed: "The capacity stated on the product page matches the physical item, so the complaint does not hold."
The draft sentence and material text one concern the same field, capacity, with mutually exclusive values. The material does not record evidence that the page statement and the physical item agree, so the draft sentence is not supported.
Model parameters are set in the application or node configuration. A compatible chat interface field of the same name cannot be assumed to override the orchestration. The same-vector-model requirement for a retrieval library does not extend to business registration websites or other external systems.
Configuration and implementation checks
The configuration below addresses the document-checking need in the exercise above. The values are hypothetical starting points for this exercise, not official defaults.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | Assume http://aiproxy:3000 | The documentation requires the root address without /v1; compare against the documented format |
AIPROXY_API_TOKEN | Assume a non-empty administrator token | The documentation requires a non-empty token matching the server; compare against the AI Proxy server configuration |
| Language model | Assume one language model is enabled | The documentation requires at least one language model; compare against the required input capability |
| Index model | Assume one index model is enabled | The documentation requires at least one index model; compare against the text-chunk indexing need |
| Model capability vision | Assume enabled as needed | The documentation states multimodal models support image recognition; compare against whether image input exists |
| Model capability toolChoice | Assume enabled as needed | The documentation states tool selection is used for classification and extraction; compare against actual node use |
Three common errors
- Entering a full request address in
AIPROXY_API_ENDPOINT. The documentation requires the root address with no/v1appended. The input to check is a hypothetical entry ofhttp://aiproxy:3000/v1; the comparison is the documented instruction not to append/v1; the observable result is that the entry does not match the documented requirement. - Using an application API key in place of
AIPROXY_API_TOKEN. The documentation requires an administrator token that is non-empty. The input to check is a hypothetical application API key; the comparison is the documented description of the AI Proxy administrator token; the observable result is that the entry does not match the required credential type. - Treating retrieval relevance as a permission filter. The documentation describes retrieval as a recall process. The input to check is a hypothetical query about a rice cooker capacity complaint; the comparison is whether permission or entity filter conditions are set; the observable result is that relevance alone cannot exclude unauthorised records.
How to verify the setup
- Check
AIPROXY_API_ENDPOINT. The input is the hypothetical entryhttp://aiproxy:3000; the comparison is the documented root-address format requirement; the observable result is that the entry contains no/v1and matches the service root address, which satisfies the condition. - Check that both a language model and an index model are enabled. The input is the two hypothetically enabled models; the comparison is the documented requirement of at least one language model and one index model; the observable result is that both model types are in an enabled state, which satisfies the condition.
- Check that the draft sentence and the material text agree on the field. The input is the draft sentence "the capacity stated on the product page matches the physical item" together with material text one, "the actual measurement is 4.2 litres"; the comparison is the value of the same field, capacity; the observable result is that the material does not record evidence that the page statement and the physical item agree, so the draft sentence is not supported and the evidence gap must be recorded.