What the documents look like
Price and display auditing for chain stores happens during day-to-day operations rather than at opening or training. The working materials come from supervision systems, POS price tables, store cameras and head-office promotion schedules, and the day-to-day documents are store visit photos, price tag lists, display standards and rectification notices. Copy drafting for products and marketing therefore handles material that carries a store, a product, a price and a promotion window. A drafted sentence can only stand when the source material states the same subject, the same period and the same field. Update frequency is weekly and monthly, and this page does not derive batch synchronization, full synchronization, real-time or transaction-triggered mechanisms from that frequency.
Constraints at the model connections step
FastGPT connects to model channels through AI Proxy. Both the App and the Admin service must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN, and a missing item causes startup validation to fail. Legacy OpenAI environment variables cannot replace these two settings. The system needs at least one language model and one index model to work normally. The model ID is the value of the model field in the request body, while the model name can be customized, and the correspondence between the two is checked when a model is enabled. Retrieval returns material; it does not grant permission over a store, a product or a price. A similarity setting does not perform entity isolation. Knowledge bases that use different vector models cannot be associated together, so identifiers are checked first and non-matching bases are excluded.
Hypothetical drill sample: the source material reads, "Store A12 in the East China region applies a shelf-edge price of 9.9 yuan from 1 March 2026 to 31 March 2026 for a 500 ml bottle of a certain brand of soda water; the head-office promotion notice for this window states that the display position for this product during this period is the second shelf of the end cap." The sentence to be judged reads, "The product was sold at 9.9 yuan for the whole of March and displayed on the second shelf of the end cap." The check is limited to the material above, and the authorization boundary is read access to this store's material for this window only.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | How it is checked and why |
|---|---|---|
| AIPROXYAPIENDPOINT | Assumed to be http://aiproxy:3000 | The official example is a service root address with no /v1 appended; compare it with the actual deployment address |
| AIPROXYAPITOKEN | Assumed to be a non-empty administrator token | The official requirement is a non-empty value consistent with the server side; compare it with the AI Proxy configuration |
| Language model and index model | Assumed to be one enabled model of each type | The official text states the system needs at least these two types; check enablement on the system model page |
| Model probe interval | Assumed to be 30 minutes | The official range is an integer between 5 and 60 minutes; compare it with the supervision rhythm before deciding |
| Knowledge base identifier | Assumed to include only this store's knowledge base for this window | Check that the vector model identifier matches; bases on different vector models are not associated together |
| Quote content limit | Assumed to follow quoteMaxToken in the selected model configuration | This field belongs to model configuration; compare it with the material volume before adjusting |
Three common errors
- Treating the model name as the model ID. The model name can be customized, while the request body uses the model ID. Mixing the two makes channel testing and actual calls point at different objects. The check is to compare the ID on the enable-model page with the model in the channel.
- Using a similarity score in place of a permission decision. A retrieval score reflects textual relevance and does not recognize store, product or price permissions. The check is to treat the authorization scope as a separate condition and not use the score as a release basis.
- Writing the period in the material as a whole month. The sample source states 1 March to 31 March, while the drafted sentence says "the whole of March", so the period wording does not match the source exactly. The check is a character-by-character comparison of the date range.
How to verify the setup
- Input the sample source material and the drafted sentence, then compare whether the price, the product specification, the display position and the date range in the drafted sentence each have a counterpart in the source with the same subject, the same period and the same field. The drafted sentence is supported by the material only when all four correspond.
- Input the model ID from the model configuration and the model name inside the channel, then compare whether the two point to the same model. A successful channel test with a recorded request duration shows the channel is usable; the test result does not show that the business copy is correct.
- Input one retrieval request and observe whether the returned results come from a knowledge base with a matching identifier. If the return is an empty array, use the paper input
[]to check how the downstream design handles an empty result, and do not infer that a rules-based retrieval necessarily returns empty.