What the documents look like
Livestream after-sales work concentrates after a single broadcast. Refund requests, reshipment orders and script templates can appear in large numbers within the same window, and agents reconcile product details against the livestream order list order by order. Product-detail and selection questions require answers that trace back to a specific document, for example whether the product name and specification in a refund request match the commitment wording in a script template. Data sources include the shop after-sales system, customer-service IM, warehouse WMS and logistics tracking, with updates after each broadcast and daily. For model connections and configuration this means documents must first be retrievable, and answers must stay inside the authorization boundary.
Constraints at the model connections step
Model access depends on AI Proxy. Under v4.17.2, both the App and any Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either item 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, and only one speech-recognition model takes effect. Retrieval is a document-recall process and does not carry permission or entity-isolation duties; a retrieval hit does not grant authorization for a refund, a reshipment or an external commitment.
Hypothetical paper exercise (not a real record): Source text one is a refund request: "Order A-001, product is a brand thermos 500 ml, buyer requests a refund, reason is a leaking lid." Source text two is a script template: "For leaking issues, apologize first and explain that a replacement lid can be sent; reshipment requires warehouse stock confirmation." Draft sentence to be judged: "Order A-001 has been approved for refund and a replacement lid has been arranged." The reviewing party is an after-sales agent, the object is order A-001, the state is refund request pending and reshipment not executed, and the effective scope is the two source texts above.
Configuration and implementation checks
Configuration centres on the model connection and reconciliation items for this page's sample. The values below are hypothetical starting points for this exercise and must be compared against the documented description or an actual observed value before being retained.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | http://aiproxy:3000 | The document requires the service root address without /v1; compare against the documented example structure and confirm no extra path. |
AIPROXY_API_TOKEN | A non-empty administrator token | The document requires it to match the server side and be non-empty; compare character by character with the AI Proxy server configuration. |
| Language model | One language model enabled | The document requires at least one language model; check isActive in the model list. |
| Index model | One index model enabled | The document requires at least one index model; check the type in the model list. |
| Model test | Run a model test on the configured channel | The document provides a channel test and a model test entry; compare the test output and request duration without assuming a pass. |
| Retrieval check | Run a retrieval with the sample source text | Observe whether the returned content contains order A-001 and the leaking-lid field; record an empty result as an empty result. |
Three common errors
- Treating a retrieval hit as authorized execution. The input is the retrieval result for the sample source text, the comparison is the refund request state and the reshipment execution record, and if retrieval returns the leaking-lid passage while the draft sentence says "approved for refund", the authorization state and the draft are inconsistent.
- Treating a model test as business verification. The input is the channel test output, the comparison is the actual handling record for the refund request and reshipment order, and a successful test only shows the model is callable, not that the order has been handled.
- Treating the script template as an executed reshipment. The input is the template wording "reshipment requires warehouse stock confirmation", the comparison is the warehouse WMS reshipment record, and the presence of the template does not mean a reshipment order was created or shipped.
How to verify the setup
- Compare the draft sentence with the source text. The input is the sample draft sentence "Order A-001 has been approved for refund and a replacement lid has been arranged"; the comparison is the refund request state "pending" and the reshipment execution record "not executed"; if the draft claims approval or arrangement, the claim is unsupported and must be changed to a pending-verification wording.
- Check the model connection configuration. The input is the actual entered values of
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN; the comparison is the documented requirement that the root address carries no/v1and the token is non-empty and matches the server side; if either item does not match, startup validation may fail, and the value must be corrected and the service restarted before rechecking. - Check retrieval against the authorization boundary. The input is a retrieval run with the sample source text; the comparison is whether the returned content contains order A-001, the leaking lid and the warehouse-confirmation requirement; if the return is empty, record it as an empty result and keep the evidence gap, without inferring that a refund or reshipment has been executed. Retrieval results serve document recall only; refund approval, reshipment execution and external commitments still require confirmation from the corresponding business system and authorization record.