Equipment Inspection and Maintenance Q&A for Semiconductor Wafer Fabrication: Multi-turn Dialogue and Prompt Engineering

Data for equipment inspection and maintenance Q&A in semiconductor wafer fabrication scenarios is typically sourced from departments including process

What this category of data looks like

Data for equipment inspection and maintenance Q&A in semiconductor wafer fabrication scenarios is typically sourced from departments including process integration, equipment, and yield improvement in wafer fabs. Core data sources include MES systems, equipment FDC monitoring data, outputs from defect detection and metrology equipment, wafer maps, process recipe libraries, and experimental data. Daily documents cover process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield reports. Data update cycles cover three dimensions: per wafer batch, real-time, and daily. Fields include nanoscale process parameters, wafer coordinates, batch ID, and other professional identifiers.

What constraints these characteristics impose on multi-turn dialogue and prompt engineering

Scattered data sources and varied update cycles in this scenario require multi-turn dialogue to distinguish between real-time equipment parameters and historical batch data to avoid context confusion. Ultra-long process chains and limited rework windows mean dialogue must retain complete historical interactions across multiple processes without losing key process information. The specificity of professional fields requires prompts to clearly specify parameter units, formats, and extraction rules, ensuring returned results comply with professional standards for wafer fabrication. Equipment abnormalities may trigger scrap of large wafer batches, so to-do items and disposal suggestions returned by dialogue must undergo data verification to prevent incorrect guidance.

How to set configurations

Configuration ItemRecommended ApproachRationale
chatIdBind to the unique identifier of the current equipment batchEnsure multi-turn dialogue is associated with historical inspection records of the same wafer batch to avoid context confusion, implemented based on FastGPT v4.17.0 session management mechanism
messagesConcatenate in the order of "system prompt + historical dialogue + current question", retain the most recent 20 rounds of interactionsAdapt to the long context requirements of multi-process wafer fabrication, cover dialogue information for the complete process chain of a single batch
variablesInject current real-time equipment FDC data and batch run record IDAllow the dialogue model to directly call on-site real-time data without additional cross-system queries
Maximum context token count8000–12000 tokensCover complete process documents and historical interactions for a single wafer batch, adapt to ultra-long dialogue flows
Output format constraintForce JSON format, exclude unnecessary line breaksAvoid JSON request failures caused by dialogue return content, comply with interface call specifications

Three common misconfigurations

  • Calling the dialogue interface returns the unAuthChat error. This occurs when the device batch chatId is not correctly bound, or the apikey is not configured with the exclusive permission scope corresponding to the wafer fab.
  • Dialogue return content includes line breaks, leading to JSON request failure. This occurs when standardized output format constraints are not injected via variables, and text returned by the model is not escaped.
  • Uploading exception handling sheets or wafer maps results in an error. This occurs when chunking rules adapted to semiconductor documents are not configured, preventing large-size SPC charts or wafer maps from being correctly parsed into dialogue context.

How to confirm configurations are properly set

  • Initiate a test dialogue containing real-time equipment FDC parameters. Verify that the model directly calls the injected variables data to return accurate inspection results.
  • Initiate three consecutive questions related to the same wafer batch. Check that chatId remains consistent and historical dialogue context is not lost.
  • Submit test text containing line breaks. Verify that returned results are correctly escaped with no JSON format parsing errors.
  • Upload an exception handling sheet and wafer map for a single batch. Check that the document is correctly chunked and included in the dialogue context with no upload errors.

The values provided are common starting points, and should be validated against operational samples specific to the deployment environment.