What Data for This Category Looks Like
The data for this category mainly comes from silicon wafer incoming material ledgers, production MES systems, operation logs of diffusion furnaces, coating equipment, and printing presses, outputs from EL and IV testing equipment, and laboratory quality analysis reports. Data update cycles follow production batches, per-wafer test nodes, and daily summaries. Daily documents include process cards, diffusion and coating parameter tables, printing and sintering records, EL detection images, IV grading datasets, and yield analysis reports. Fields cover process parameter values, electrical performance indicators, appearance defect markers, and more.
Constraints Imposed on Model Access and Configuration
Multi-source heterogeneous data types require model access to adapt to structured parameters, time-series logs, and image inputs simultaneously. Frequently updated production data requires model call latency to align with production node rhythms, preventing delayed tender data. Large single-batch data volumes with subdivided fields require limiting the context length of a single input to avoid exceeding model thresholds. Tender documents need to integrate multi-process data, so models must support cross-document associated extraction, and adapt to concurrent scenarios for batch data processing.
Configuration Settings
These configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
vision | Enabled | Required to process visual data such as EL detection images and IV test images, to support defect recognition and indicator extraction |
maxContext | 8000–16000 token | Adapts to the total text length of single-batch production data and tender documents, to avoid context overflow |
maxTokens | 2000–3000 token | Controls the output length of a single response, to ensure quotation and description content in tender documents is complete and concise |
quoteMaxToken | 1500 token | Limits the length of referenced source document content, to prevent redundant information from interfering with the core logic of tender documents |
temperature | 0.3–0.5 | Maintains the accuracy of quotations and process parameters, avoiding generated content that deviates from actual data |
AIPROXY_API_ENDPOINT | Fill in according to deployment environment | Adapts to local or cloud model service addresses, to ensure normal data transmission |
AIPROXY_API_TOKEN | Key applied for according to permissions | Ensures identity legitimacy for model calls, preventing unauthorized access |
Three Common Misconfigurations
- Uploading EL detection images triggers visual model call errors, returning format unsupported or 400 errors. Cause: The
visionconfiguration item is not enabled, or the model backend adapted for visual tasks is not specified, resulting in failure to parse image formats. - No response occurs when shared page concurrency exceeds 60, while local model server resources have not reached their bottlenecks. Cause: The resource threshold for concurrent calls is not properly configured, or request queue current limiting is not set, resulting in model service overload.
- Uploading large documents containing IV grading data returns a 400 error, or there is data within the <FilesContent> tag but the model fails to extract valid content. Cause: The
maxContextlimit is not set, and the total document length of a single input exceeds the context upper limit supported by the model; or long documents are not preprocessed and split, resulting in the model being unable to read content completely.
How to Confirm Proper Configuration
- Upload a single EL detection image and the corresponding process card document, verify that the model can correctly recognize defect markers in the image and associate process parameters.
- Simulate multiple concurrent requests, observe the response status of the model service, and adjust the configuration to match actual concurrency requirements.
- Upload documents exceeding the conventional single-batch data volume, check whether context overflow errors are triggered, and confirm that the
maxContextconfiguration is reasonable. - View model call logs, verify that the
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENconfigurations have completed identity verification normally.
The values provided here are common starting points, and should be measured against the reader's own samples.