What the documents look like
Seller ratings management for platform stores works from consumer feedback that arrives after fulfillment. The daily materials are negative-review handling records, experience-score detail sheets, service diagnostic reports and remediation ledgers. Typical roles are store operations staff, customer service supervisors and quality inspection specialists. When these materials feed product-information and selection questions, the model-connection step must first settle which data source and which document boundary applies. Experience-score detail sheets and service diagnostic reports are commonly refreshed daily or weekly; negative-review handling records appear as each after-sales case progresses. That refresh cadence does not change the connection configuration on this page. It changes only the point in time at which material versions are checked during handover.
Two industry facts shape this task. First, the feedback data comes from post-fulfillment sources such as store experience scores, platform service indicators, after-sales reviews and the customer service system, so a question about product information and selection is answered from records that describe service outcomes rather than from promotional or pricing material. Second, the most time-consuming work is attributing a negative review, contacting the buyer for supporting evidence, and tracking the indicator after remediation. A draft sentence that treats a pending remediation as finished therefore fails against the same field in the same record, and the model-connection step has to keep that distinction visible rather than smoothing it away.
Constraints at the model connections step
FastGPT connects to model providers through the AI Proxy service. From V4.17.0, AI Proxy is a required dependency. Both the App and an Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail. Legacy OpenAI environment variables or a single-model custom request address cannot substitute for these two settings. The system needs at least one language model and one index model to work normally. Only one speech recognition model takes effect, so configuring one is sufficient.
Product-information and selection questions are an application-layer orchestration task. Model capability is checked against what the application actually needs as input. Multimodal support, tool choice and thinking configuration are verified separately. Application parameters are set in the application or node configuration; there is no basis to assume that a same-named field on a compatible chat interface overrides the orchestration. A thinking budget is a reasoning budget and cannot be moved to serve as a quote token ceiling. Retrieval and relevance scores do not carry permission or entity isolation, and the same-vector-model requirement for a linked knowledge base does not extend to business-registration sites or other external systems. Knowledge bases built on different vector models cannot be associated together; the identifier is checked first and non-matching items are excluded, with the actual observation recorded.
Configuration and implementation checks
The following items concern this page's sample and the named task. They list documented settings or implementation checks only. Hypothetical trial sample: a store operations staff member on 10 March 2026 compiles a negative-review handling record. The source text reads: "Order A was signed for on 5 March. The buyer reported on 6 March that the item description did not match the goods. Customer service contacted the buyer on 7 March to request supporting evidence. As of 10 March the evidence has not been received, and the remediation ledger status is pending verification." The draft sentence to be judged is: "This order has completed remediation and been closed." The comparison material is the service diagnostic report for the same period in the experience-score detail sheet.
| Configuration item | Hypothetical starting point or check content for this run | Compared against |
|---|---|---|
| AIPROXYAPIENDPOINT | Hypothetical entry of the AI Proxy service root address, with no /v1 appended | The actual deployed root address |
| AIPROXYAPITOKEN | Hypothetical non-empty administrator token | Server-side configuration, matching and non-empty |
| Language model | Hypothetical enabling of one tested language model | The requirement for at least one language model |
| Index model | Hypothetical enabling of one tested index model | The requirement for at least one index model |
| Model capability | Hypothetical check of multimodal and tool choice against application input needs | The model configuration fields vision and toolChoice |
| Application parameters | Hypothetical setting in the application or node configuration | The parameters the orchestration actually reads |
Reason for the starting points: each value above is a fill-in for this hypothetical trial, used to check whether the configuration is complete. None is an official default.
Three common errors
- Treating a pending-verification status as completed remediation. The input is the sample source text and the draft sentence. The comparison is the value of the same field in the remediation ledger. The observable result is that the draft sentence says remediation is complete and the order is closed, while the source text records the evidence as not received and the ledger status as pending verification.
- Treating a retrieval result as authorized execution. In the sample, the material is used only for product-information and selection questions. A retrieval hit on the negative-review handling record does not mean authorization has been obtained to contact the buyer or close the order. The input is the retrieval request and the material scope; the comparison is the authorization-boundary statement; the observable result is that the material states no authorization clause.
- Entering a full model request address in
AIPROXY_API_ENDPOINT. The documentation describes this variable as the service root address, with no /v1 appended. The input is the configured value; the comparison is the documented requirement; the observable result is that the value ends with /v1 or /chat/completions.
How to verify the setup
- For the sample source text and the draft sentence, the input is "Order A was signed for on 5 March ... the remediation ledger status is pending verification" together with the draft sentence "This order has completed remediation and been closed." Compare the remediation ledger status value for the same order and the same field. The condition is met when the draft sentence's status matches the pending-verification status in the source text. When it does not match, the draft sentence is unsupported.
- For the authorization boundary, the input is the sample material and the retrieval request. Compare whether the material states an authorization clause for contacting the buyer or closing the order. The condition is met when the material clearly states the authorizing party and scope. When the material states none, record an evidence gap.
- For the
AIPROXY_API_ENDPOINTconfiguration, the input is the hypothetically entered service root address. Compare whether that address appends /v1 or /chat/completions. The condition is met when the address is only the service root address. When a path is appended, the entry does not match the documented requirement. All three are paper checks; no live test is claimed.