This document is written based on FastGPT v4.17.0.
What this category’s data looks like
Daily documents for this category include recipe sheets, weighing records, mixing curves, rubber compound quick tests, vulcanization process cards, mold records, and finished product performance reports. Data sources cover recipe systems, raw material batch ledgers, internal mixers, Mooney viscometers, rheometers, vulcanizing presses, and laboratory testing equipment. Data is updated per raw material batch, per mixed rubber batch, and per vulcanized product mold run. Single records include fields such as recipe ratio, process temperature, duration, batch number, and test values, with units including degrees Celsius, minutes, kilograms, and Shore hardness, among others.
Constraints on model access and configuration
The high-frequency, multi-source heterogeneous data characteristics of this category create multiple constraints for model access and configuration. Real-time updated multi-batch data requires the context length to match the data volume, to avoid truncating key audit basis. Image documents such as mixing curves and vulcanization process cards require models to support visual recognition capabilities. Models deployed on the internal network must operate in the same network environment as the FastGPT server, and cannot rely on public external network APIs. The precise verification requirements for multiple fields demand stable model output, with excessive randomness to be avoided. In addition, the response time of a single task must align with production rhythm, to prevent delays that affect on-site audits.
How to set configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Covers historical records and current audit data for multi-batch mixing and vulcanization to avoid context truncation |
maxTokens | 2000–3000 token | Adapts to parameter verification and deviation judgment needs for single-batch audits, controls response time to match production rhythm |
temperature | 0.1–0.3 | Ensures judgment stability for quality audits, reduces the impact of random output on audit results |
vision | Enabled | Supports recognition of parameters and records in image documents such as mixing curves and vulcanization process cards |
AIPROXY_API_ENDPOINT | Internal network deployed DeepSeek API address | Adapts to the network environment where both the FastGPT server and the model are on the internal network, and cannot access external network resources |
| Workflow/API tool selection | auto | Automatically calls tools to extract quality parameters from multi-source data, no need to manually specify call logic |
Three common misconfigurations
- Configuring
AIPROXY_API_ENDPOINTas a public external network address will return a connection timeout error during calls. This is because the FastGPT server is located on the internal network and cannot access external network resources. - Setting
maxTokensto5000 tokenwill frequently cause response truncation in audit tasks. This is because the fields and historical data that need to be covered for single-batch audits in this category are numerous, exceeding the preset token limit. - Not enabling the
visionparameter will prevent recognition of temperature fluctuation records in mixing curve images, leading to audit results that miss process deviations.
How to confirm proper configuration
- Upload a mixing curve image document, trigger an audit task, and verify that core parameters including temperature and time in the image are accurately extracted.
- Submit an audit request containing multi-batch mixing and vulcanization data, and check that no context truncation prompt appears in the response content.
- Directly call the configured API address in the internal network environment, receive a normal model response, and confirm that the authentication token is valid.
- Adjust the
temperatureparameter to0.5, compare the results of two identical audit tasks, and confirm that the parameter configuration takes effect.
The values provided are common starting points and should be measured against the reader's own samples.