What Data Looks Like for This Category
Daily documents include batching lists, smelting records, electrolytic cell condition records, laboratory reports, anode and cathode quality records, and environmental protection ledgers. Data sources cover concentrate and auxiliary material analysis, DCS systems, cell voltage and current monitoring, laboratory data, and energy and environmental protection monitoring platforms. Update rhythms fall into three categories: per batch of raw materials, per furnace or per cell, and real-time. Document fields include grade, moisture content, feeding amount, furnace temperature, cell voltage, current value, and other metrics. Corresponding units are mostly percentage, ton, degree Celsius, volt, and ampere.
Constraints on Model Access and Configuration
Multi-source heterogeneous data sources create differences in field formats and units across documents. Model access must adapt to align multi-source data, so context stitching parameter limits must be adjusted. Real-time cell condition data requires model response speed to match production rhythms. An overly large context window must not be set. Differences in update frequencies across documents require configuration of flexible context refresh rules. This avoids redundant data occupying model resources. Compliance-focused environmental protection ledger documents require fixed output formats. The model’s free generation space must be limited to ensure review results meet business specifications.
Configuration Settings
This configuration is adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to field stitching for multi-source documents, avoids exceeding model context limits, and covers all core documents for a single production batch |
maxTokens | 2000–3000 token | Matches the review output length of a single technical document, balances response speed and content completeness, and adapts to production rhythm requirements |
quoteMaxToken | 1500 token | Limits the token proportion of referenced source documents, ensuring review logic focuses on core business fields such as grade and cell voltage |
temperature | 0.2–0.4 | Ensures consistency of review results, avoids deviations in impurity accounting and metal recovery rate calculations caused by parameter fluctuations |
| Workflow/API tool selection | auto | Automatically selects whether to call tools to handle multi-source data alignment, adapting to format differences across different document types |
AIPROXY_API_ENDPOINT | Configured according to enterprise intranet deployment nodes | Adapts to the model access path of the enterprise intranet, ensuring transmission security of sensitive production data |
Three Common Misconfigurations
- Phenomenon: Abnormal runtime statistics are returned by model calls. Offline models display total time spent, while online models only display first-token time spent. Unified viewing is not possible. Cause: Response parsing rules for
AIPROXY_API_ENDPOINTare not uniformly configured, and return formats of different models are not adapted. - Phenomenon: An API Key error prompt is returned when calling a workflow. The prompt indicates an incorrect key type was used. Cause: Global universal API Key and application-specific key are mixed, and keys with the required permissions are not configured as specified.
- Phenomenon: Some core fields such as cell voltage and grade are missing from review results. Cause: The
maxContextparameter is not adjusted, leading to multi-source document stitching exceeding context limits, resulting in truncation of some fields.
How to Confirm Successful Configuration
- Upload a single batching list and electrolytic cell condition record. Check whether the review results returned by the model include all core business fields, with no missing or truncated content.
- Call a workflow test. Check whether the returned runtime statistics uniformly display total time spent, with no abnormal situation where only first-token time is displayed.
- Verify API Key configuration. Use an application-specific key to initiate a call, and confirm no permission error prompts are displayed.
- Adjust the
temperatureparameter to 0.3. Upload the same document multiple times, and check whether result fluctuations fall within an acceptable range, with no obvious deviations.
The values provided are common starting points and should be evaluated against relevant on-site samples.