What the data for this category looks like
Data sources include Gerber files, drill files, and BOM lists provided by clients. Additional sources are engineering documents generated by CAM, production equipment parameters, bath solution test results, AOI inspection reports, electrical testing reports, and slice analysis reports. Data is collected and updated continuously for each production batch and shift. Single-batch documents contain engineering production instructions, drilling and lamination parameters, chemical analysis records, and other content. Fields include layer specifications, processing parameters, test results, and related items. Units used include millimeters for aperture, degrees Celsius for temperature, megapascals for pressure, and other standard units.
Key Constraints for Multi-turn Dialogue and Prompt Engineering
The multi-source, heterogeneous and frequently updated document characteristics of this category require multi-turn dialogue to associate different types of production and test data. This avoids bias from single-source information. Documents updated per production batch require dialogue context to support filtering data by production batch. Documents include parameters with attached units. Prompts must explicitly specify fields and units to prevent parameter confusion. Exception report drafting requires association with multi-dimensional test results. Multi-turn dialogue must trace the root cause of exceptions step by step. Single-round generation cannot cover all relevant information.
Configuration Setup
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chatId | Bind production line number and production batch number | Enable independent context management for multi-batch conversations on a single production line |
messages | Retain the last 10 conversation history entries | Limit context redundancy and focus on the current batch’s conversation progress |
variables | Bind production batch filtering rules | Quickly call documents for the target batch, such as Gerber files and AOI reports |
| Maximum context length | 8000–10000 characters | Match the total character count of single-batch production documents, avoiding truncation of critical parameters |
| File upload storage path | Set based on actual testing | Adapt to storage configuration requirements for private deployments |
Three Common Misconfigurations
- Symptom: After enabling the file upload function in a workflow, calling the document parsing tool returns an "invalid file path" error. Cause: The storage path for uploaded files is not configured correctly. This prevents the parsing tool from accessing uploaded Gerber files or test reports.
- Symptom: Responses generated via code calls to the dialogue interface do not associate with production documents in the knowledge base. Cause: The production batch identifier for context association is not correctly set in the
messagesparameter. This causes the model to not retrieve engineering documents for the target batch. - Symptom: The AI dialogue component triggers an "uncaught exception" prompt in a private deployment environment. Cause: The production batch filtering rule bound to the
variablesconfiguration is not correctly set in the configuration file. This prevents the model from matching the target data.
How to Verify Correct Configuration
- Upload a single batch’s Gerber file and AOI report, initiate a conversation, and check if the response references parameter information from the uploaded files.
- Configure
chatIdto bind production line numbers, initiate conversations for two different production lines, and confirm that contexts do not interfere with each other. - Configure
variablesto bind production batch rules, initiate a conversation, and check if the model automatically filters production data for the target batch. - Run a workflow test for the file upload and parsing process, confirm no errors occur, and verify that the parsed results include core fields from the documents.
The values provided here are common starting points. They should be evaluated against the user’s own samples.