Livestream violation appeals: model connections for product questions

What the documents look like

Livestream violation appeals respond to penalties that have already been issued. The material set is organized around the violation notice, the livestream recording, product qualification files, and the appeal statement. The product information and selection question-answering task turns the checkable entries in product qualifications and appeal statements into a searchable, quotable question-answering basis, so that compliance and operations staff can locate segments when drafting appeal reasons. Material arrives irregularly; each violation triggers one round of collection. The ingestion scope, model capabilities, and citation boundaries therefore need to be agreed in advance, so that material not ingested is not treated as verified evidence.

The practical consequence for model connections is that the question-answering layer sits downstream of a penalty that already exists. A retrieval hit is a recall result, not a permission or entity filter, and retrieval does not constitute authorization. A similarity score does not exclude unrelated businesses or unauthorized records.

Constraints at the model connections step

Three constraints shape this step. First, AI Proxy is a required dependency from V4.17.0. The App service and any Admin service that reuses App/Service configuration must both set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN; missing either item causes startup validation to fail. Legacy OpenAI environment variables or a single-model custom request address cannot replace these two settings. Second, the system needs at least one language model and one index model to work normally, and only one speech recognition model takes effect. Third, the model ID is the value of the model field in the interface request body and is globally unique, while the model name is for display only. Model mapping maps the model FastGPT requests to the upstream model actually provided, for example mapping gpt-4o-test to gpt-4o.

Hypothetical paper exercise sample: a shop receives a platform violation notice stating that during a presentation on 12 March 2025 from 20:00 to 21:00, a claim of "keeps warm for 24 hours" was made about a vacuum flask, while the test report for that flask in the product qualification library records a thermal retention performance of 12 hours. The draft sentence to be judged is: "The warm-keeping duration described in this presentation is consistent with the product qualification record." This sample is for paper comparison only; nothing was submitted, nothing was retrieved, and no approval was obtained.

Configuration and implementation checks

The table lists only settings documented in the official material or implementation checks tied to this page's main task. Values marked as a hypothetical starting point are starting values for this paper exercise; the reason for choosing them and the value they are compared against appear in the description column. They are not official default values.

Three common errors

1. Writing AIPROXYAPIENDPOINT as a URL that contains /v1 or a complete chat address. Check: take the value of that variable, compare it against the documented statement not to append /v1, and treat any extra path as not satisfying the condition.
2. Using the application API Key as AIPROXYAPITOKEN. Check: take that token and the server-side administrator token, compare whether they match and are non-empty, and treat a mismatch or an empty value as not satisfying the condition.
3. Treating a retrieval hit as verified evidence. Check: take the draft sentence and the returned segment, compare whether the segment comes from the product qualification and whether the field is thermal retention performance, and treat a similarity score alone as not constituting verification.

How to verify the setup

• Verification one: take the product qualification text from the hypothetical sample, "the test report records a thermal retention performance of 12 hours", and the draft sentence "the warm-keeping duration described in this presentation is consistent with the product qualification record". Compare the values of the two on the same field. A value of 12 hours and a value of 24 hours are mutually exclusive values on the same field, so the draft sentence is not supported. If the source text does not state thermal retention performance, record the item as pending verification rather than as a negation.
• Verification two: take the handover request "please submit the above draft sentence to the platform appeal". Compare whether the request carries a signed appeal statement and an authorization record. No signature or authorization material appears in the sample, so the request remains in a pending-authorization state and does not constitute an executed action.
• Verification three: take the paper input of an empty retrieval result [] and check whether the downstream design outputs "no basis found" when no segment is present. If the downstream directly generates a conclusion sentence, that does not match the requirement to handle an empty result; this empty result is a paper input only and does not indicate that any real retrieval returns an empty value.

All of the above is paper comparison. No model was invoked, no retrieval was performed, no appeal was submitted, and no approval was obtained.

Related reading

• self-host/config/env
• self-host/config/model/intro
• FastGPT model connections