What the documents look like
Channel diversion review for wholesale and distributor channels handles regional and price order. It does not handle rebate amounts, which belong to rebate accounting. The working materials are diversion complaint forms, anti-counterfeit code records, penalty notices and investigation photographs. A single shipment's anti-counterfeit code records may originate from different regions, and when complaint forms and photographs lack a shared identifier, reconciliation stays manual. Materials update weekly or irregularly; this page does not derive batch synchronization, full synchronization, real-time or transaction-triggered mechanisms from that cadence.
The product information and selection question-answering task depends on a language model and an index model. The system needs at least one language model and one index model to work normally. Multimodal capability decides whether investigation photographs can enter question answering, and tool-calling capability affects classification and content extraction. These capabilities are checked separately against the inputs an application actually needs; a same-named field on a chat interface does not establish orchestration behavior. Retrieval recalls text chunks and does not perform authorization decisions. A similarity setting does not exclude unauthorized records.
Constraints at the model connections step
The step concerns model access and configuration for product information and selection question answering, not the penalty workflow itself. FastGPT reaches model providers through AI Proxy. Since V4.17.0, AI Proxy is a required dependency: App and Admin services that reuse App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN, and a missing item causes startup validation failure. Legacy OpenAI environment variables and single-model custom request addresses do not replace these two settings. Model ID is the value of the model field in a request body and is globally unique. Model name is a display label. Model channel is a provider protocol, and one model can be configured across channels for load balancing. Model mapping changes the model value sent upstream, as in mapping gpt-4o-test to gpt-4o; the in-system model ID is unchanged. Custom request address and key bypass the model channel and are generally not recommended.
Hypothetical drill sample: a brand channel-control team receives one diversion complaint form whose original text reads: "Complaint target: East China regional distributor A; product involved: one model of goods; complaint reason: a listing below the suggested retail price found in a South China e-commerce shop; attachments: two investigation photographs and one anti-counterfeit code record; complaint form status is pending review." A separate draft sentence awaits judgment: "Distributor A has completed cross-region sales and the penalty notice has been issued." In the reconciliation materials the complaint form status is pending review, and no penalty notice text appears. The penalty fact in the draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reconciliation method |
|---|---|---|
| AIPROXYAPIENDPOINT | Hypothetically filled with the AI Proxy service root address, without /v1 appended | Compare character by character against the documented root-address format |
| AIPROXYAPITOKEN | Hypothetically filled with a non-empty administrator token matching the server side | Confirm non-empty; confirm an application API key was not substituted |
| Language model | Hypothetically one language model enabled | Confirm enabled; confirm the model ID matches the model in the channel |
| Index model | Hypothetically one index model enabled | Confirm enabled; confirm the model ID matches the model in the channel |
| Model capability | Hypothetically checked against the text input this sample needs | Check multimodal and tool-calling separately; do not merge the judgments |
| Model status probing | Hypothetically automatic probing not enabled | When not enabled, the page presents the automatic probing not enabled empty state |
The starting values are hypothetical trial settings for this drill, used to compare against documented formats or states. They are not official default values.
Three common errors
- Treating the complaint form's pending review status as confirmed diversion. The input is the sample complaint form text; the comparison object is the complaint form status field. Pending review does not constitute a confirmed conclusion, and the penalty fact in the draft sentence is not supported.
- Treating a retrieval hit as authorization approval. The input is one product information retrieval request; the comparison object is the retrieval result against access permission. A retrieval result does not establish that the account may view the corresponding record.
- Writing model mapping as model replacement. The input is the model mapping in the channel configuration; the comparison object is the FastGPT request model against the actual upstream model. Mapping changes only the model value of the upstream request and does not change the in-system model ID.
How to verify the setup
- Input the sample complaint form text and the draft sentence; compare whether the draft sentence contains the penalty notice having been issued. The condition is met only when penalty notice text appears in the materials and the status supports it; otherwise record it as unsupported.
- Input the sample complaint form text and a question asking whether distributor A engaged in cross-region sales; compare whether the answer cites the pending review status. An answer that states pending review as confirmed does not meet the condition.
- Input the paper input
[]for the empty-result downstream scenario; compare whether the downstream design handles it as an empty result. Do not claim that a fictional rule retrieval necessarily returns[]; record observed return content and error information as they are, and leave unobserved parts unknown.
For the model status probing described in the official documents, only enabled system models are probed; a first failure is retried up to three more times, and only four failed attempts in one round are finally judged abnormal. Normal, high-latency and abnormal states correspond to green, yellow and red. Webhook pushes occur on edge changes, with event types model_status_error and model_status_recovered, and a storm-prevention mechanism suppresses repeated pushes while a model stays abnormal. This page does not assert that probing was enabled, that any webhook fired, or that any model was tested in this scenario.