What this category of data looks like
Daily documents for this category include sales orders, master production schedules, material requirements plans, daily and weekly production schedules, work orders, material shortage lists, and delivery change notices. Data is sourced from ERP orders and inventory, bills of material, process routes, equipment capacity, staff shifts, work in progress, and supplier deliveries. Updates occur daily, weekly, or when triggered by orders or exceptional events. Document fields include order numbers, material codes, capacity units, process durations, work in progress quantities, and more. There is no unified fixed format, and configurations vary based on individual enterprise ERP systems.
Constraints on model access and configuration
Data sources for this category are scattered, and update rhythms are flexible. Model access configuration must adapt to multi-format input and dynamic data loading. Multi-source heterogeneous document structures increase model context processing pressure. Reasonable limits must be placed on the scale of content input in a single session. Quotation and scheduling results in bidding scenarios require high consistency, which creates clear requirements for model output stability. Data update frequency is also high, so configuration must support rapid synchronization of the latest order and capacity information. This prevents scheduling deviations caused by outdated data. Differences in document fields across enterprises add a need for configuration flexibility to adapt to varying input formats.
How to set configurations
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
maxContext | 8000–12000 token | This category of data includes multiple related documents. Sufficient context capacity is needed to associate order, scheduling, and capacity data |
maxTokens | 4000 token | Limit single output length to avoid generated bid documents and scheduling content exceeding formatting specifications |
temperature | 0.2–0.4 | Reduce output randomness to ensure consistency of scheduling parameters and quotation logic, and meet the rigor requirements of bidding scenarios |
AIPROXY_API_ENDPOINT | Fill in the locally deployed model API address | Adapt to production planning-related models deployed within the enterprise, and ensure data stays within the enterprise intranet |
AIPROXY_API_TOKEN | Fill in the actual generated secret key | Verify model call permissions and prevent unauthorized access |
Three common mistakes
- Symptom: Call to a locally deployed m3e model fails, with connection timeout shown in logs. Cause: In FastGPT v4.17.0, the
AIPROXY_API_ENDPOINTconfiguration does not correctly point to the local model's OpenAPI port, or port permissions are not enabled. - Symptom: Error occurs when connecting to a oneapi embedding model, and the index model cannot be configured separately. Cause: The configuration entry for chat models and index models is not correctly distinguished, and the embedding model was mistakenly added to the chat model parameter field.
- Symptom: Scheduling result fields returned after model call are missing, such as process duration and material code not being generated normally. Cause: The
maxContextvalue is too small, failing to fully load associated BOM and process route data, which prevents the model from extracting necessary fields.
How to confirm successful configuration
- Run a model call test, input a simulated sales order and capacity data, and check that the returned scheduling results include core fields such as processes, materials, and delivery dates.
- Check system logs to confirm there are no permission or connection errors related to
AIPROXY_API_ENDPOINTorAIPROXY_API_TOKEN. - Import multiple different types of daily documents, such as work orders and material shortage lists, to verify that the model can correctly associate information from different data sources.
- Adjust the
temperatureparameter to 0.3, call the same input three times consecutively, and confirm that the difference in output results is within an acceptable range.
The values provided here are common starting points, and should be evaluated against the reader's own samples.