What the documents look like
Warehouse picking, review and packing handles order outbound work rather than supplier arrivals. Its operational records come from the WMS, the order system, the electronic waybill platform and weighing equipment, and they change per order or per hour. The daily documents in this step are wave picking lists, review records, waybills and packaging consumable sheets. Copy drafting for goods and marketing draws on the text of these documents, then checks whether a proposed sentence is supported by that text. Because each order produces its own record, the same item can carry different review entries under different waves, so any draft must state the wave, date and document scope it relies on.
Constraints at the multi-turn prompts step
Multi-turn dialogue distinguishes a new session from a continued session. When a chatId is passed, FastGPT takes the session from its database and uses the last element of the messages array as the user question, ignoring the remaining messages. When no chatId is passed, the context is built entirely from the supplied messages. The prompt must name the material scope, the answering task and the handling of missing evidence, so the model does not fill in business facts on its own. Memory turns control how many earlier rounds of chat the answer may reference; a larger value reaches further back and carries more content. Reply limit, temperature, Topp and reply format follow the downstream requirement. Selecting jsonschema also requires filling in the corresponding format requirement and checking the returned structure.
Hypothetical drill sample: the source material reads "Wave picking list W-0417, 17 April 2025, item A picked quantity 20 units; review record R-0417 states 20 units actually reviewed, status reviewed." The draft sentence under judgment reads "Item A completed outbound review of 20 units on 17 April 2025." The subject, object, date and quantity can be compared against the source, while whether "completed outbound" equals "reviewed" is an item to verify; the source does not state an outbound action.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| AI model | Select a model that supports multimodal recognition | Used when reading a waybill or item image; the popup shows whether the model supports the relevant media type, compared with a model that does not |
| Memory turns | A small number of turns for now | Copy drafting usually needs only the current wave material; compared with a larger number, observe whether unrelated history is carried in |
| Temperature | Set low for now | The draft must stay close to the source text; compared with a higher value, observe whether the draft sentence departs from source facts |
| Reply limit | A medium length for now | A complete copy paragraph is required; compared with a very low value, observe whether output is truncated early |
| Reply format | Keep the default first | No downstream fixed-structure read is required; if the downstream needs a fixed format, switch to json_schema and fill in the format requirement |
| Multimodal recognition | Enable image recognition according to model capability | Used when the material contains a waybill or item image; compared with the off state, observe whether the model reads the image content |
Three common errors
- Writing "reviewed" from the review record directly as "completed outbound." The input is the source material above and the draft sentence; compare whether "reviewed" and "completed outbound" are the same field with the same value; the observable result is that the source does not state an outbound action, so the sentence is unsupported.
- Omitting chatId while expecting the system to bring in prior session history automatically. The input is a request without chatId; compare the messages actually carried against the expected context; the observable result is that the context is built entirely from the supplied messages and history is not taken automatically.
- Selecting jsonschema without filling in the corresponding format requirement. The input is a configuration that selects jsonschema; compare whether the format requirement is filled in and what structure is returned; the observable result is that without the format requirement the fixed structure cannot be returned.
How to verify the setup
- Use the hypothetical source material and draft sentence above. The input is the source material, the draft sentence and the prompt; compare the subject, date and quantity in the draft sentence against the corresponding fields in the source; the observable result is that date and quantity match while "completed outbound" is not supported by the source and must be marked for verification.
- Use a request without chatId. The input is a messages array containing only the current question; compare the context actually carried against the supplied messages; the observable result is that the context consists only of the supplied messages. If chatId is passed, compare the session the system takes against the last element of messages and observe whether they agree.
- Use the hypothetical empty-result handoff input []. Compare whether the downstream design handles an empty reference; the observable result is that the downstream follows an empty-reference branch. This check is a paper design check and does not state that a fictional rule retrieval necessarily returns an empty array.