What data looks like for this category
Daily documents for this category include formulation sheets, weighing records, mixing curves, rubber compound quick tests, vulcanization process cards, mold records, and finished product performance reports. Data sources cover formulation systems, raw material batch management, internal mixers, Mooney viscometers and rheometers, vulcanizers, and laboratory testing equipment. Data updates occur once per rubber batch, per raw material batch, or per molded product. Single-batch data includes structured fields such as feeding ratios, process parameters, and test results. Field units vary by document type. For example, mixing curves include temperature, rotational speed, and duration parameters. Vulcanization process cards include pressure, temperature, and curing duration parameters.
What constraints these characteristics impose on model access and configuration
High-frequency single-batch data updates require model access to adapt to small-batch, high-frequency document indexing and retrieval workflows. This avoids production delays caused by batch processing lags. Multi-source heterogeneous device data and laboratory test data coexist. Configuration must define clear field mapping rules for different data sources. This ensures the model can integrate correlated data across systems. Many structured fields have strong business correlation. For example, mixing curves correspond to the same batch’s vulcanization process card. Configuration must retain contextually linked parameter settings. This prevents logical breaks in generated content caused by splitting correlated data. Process and exception reports require strict accuracy. Configuration must constrain the model’s generation flexibility.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | The total token count of single-batch linked documents (mixing curve + vulcanization process card + test report) for this category usually falls within this range, covering contextual information for the complete process chain |
temperature | 0.2–0.4 | Process documents and exception reports require strict accuracy. This parameter range reduces randomness in generated content, avoiding deviations from actual production parameters |
quoteMaxToken | 1500–2000 characters | Full reference of key single-batch process parameters must be retained. This range avoids truncation of core production data |
| Workflow/API tool selection | auto | Automatically select whether to call retrieval tools to obtain latest batch linked data based on input content, adapting to high-frequency update business scenarios |
AIPROXY_API_ENDPOINT | Exclusive API address for locally deployed models | Meets the requirement of this category’s enterprises often deploying local models to ensure production data security |
AIPROXY_API_TOKEN | Access key generated per enterprise internal security rules | Ensures permission security for model calls, preventing unauthorized access to production data |
Three common configuration mistakes
- Issue: After deploying the embedding-2 or m3e vector model locally, configuring
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENcontinuously shows "indexing" with no progress updates. Cause: The model's API response format is not adapted to FastGPT's vector retrieval requirements, so the platform cannot correctly parse the returned vector data. - Issue: After accessing a locally deployed deepseek-r1 model, the generated report contains
<think></think>tag content. Cause: The native thought process output is not disabled in the model configuration, or label replacement rules that meet FastGPT requirements are not configured. - Issue: After configuring speech recognition and reranking models, the generated process documents have missing fields or abnormal values. Cause: Output format constraints for the model are not configured for the structured document fields of this category, such as mixing temperature and curing duration, leading to generated content that does not meet business field requirements.
How to confirm the configuration is correct
- Upload a single batch of mixing curves and vulcanization process cards, trigger the model to generate a process document, and check if the generated content includes the key process parameters of the corresponding batch.
- Simulate an exception scenario such as under-curing or over-curing, input the test data of the corresponding batch, and check if the model can correctly generate an exception report and link to the linked data of historical batches.
- View the model call logs in FastGPT v4.17.0, confirm that the actual effective values of configuration parameters such as
maxContextandtemperaturematch the preset values. - Test the API call of the locally deployed model, confirm that the return format meets FastGPT's access requirements, with no format errors or missing fields.
The values provided are common starting points and should be measured against the reader's own samples.