What This Category of Data Looks Like
Data generated by engineering, production, quality control, and laboratory teams at printed circuit board factories includes customer-provided Gerber files, drill files, BOMs, internal CAM output files, equipment operating parameters, bath solution test results, AOI inspection reports, electrical test reports, and slice analysis records. Daily documents cover engineering production instructions, drilling and lamination parameter logs, chemical analysis reports, and various types of inspection reports. Data update frequency follows per production batch, per shift, or continuous collection. It covers the full workflow from customer data conversion to finished product inspection. Fields and units vary by document type, with no unified fixed standard.
Constraints for Knowledge Base Retrieval and Recall
Scattered data sources and diverse formats require the retrieval system to support multi-format data parsing, including both structured process parameter documents and unstructured inspection reports. High update frequency and tight binding to production batches demand precise matching of batch identifiers to avoid cross-batch data confusion. Single-batch data volume fluctuates with production scale, so the number of retrieval results must be controlled to prevent context overload that interferes with audit judgments. Core workflow focuses on engineering data conversion and electrical connectivity, so retrieval must prioritize associating key process documents such as Gerber parsing and drilling parameters to support the accuracy of inspection and audit.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the precision requirements for PCB process parameters and inspection reports, avoids irrelevant low-matching data interfering with audits |
limit | Top 10 entries | Controls the number of documents returned per batch retrieval to prevent context overflow that affects audit logic |
searchMode | fullTextRecall | Adapts to mixed retrieval needs for multi-format documents, covering structured parameters and unstructured report content |
usingReRank | Enabled | Performs secondary sorting on initial retrieval results, prioritizes returning high-value documents related to process parameters and electrical connectivity |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands keywords such as batch identifiers and Gerber numbers based on user queries to resolve reference resolution issues |
collectionIds | Knowledge base collections bound to production batches | Precise limitation of retrieval scope, avoids cross-batch data confusion, improves audit targeting |
Three Common Configuration Mistakes
- Uploading HTML-format interface documents results in no retrieval returns. Cause: FastGPT's built-in HTML document parsing plugin is not enabled, or the appropriate parsing mode is not selected during upload.
- Retrieval results do not associate production batch information, leading to cross-batch process data confusion. Cause: The exclusive knowledge base collection for the current production batch is not bound via
collectionIds, anddatasetSearchUsingExtensionQueryfor automatically expanding batch keywords is not enabled. - GPU resources are not occupied after triggering knowledge base retrieval. Cause: GPU acceleration parameters for the vector database are not configured in FastGPT v4.17.0, or the vector database background service is not running properly.
How to Confirm Configuration Is Complete
- A test PCB drilling parameter HTML document is uploaded, retrieval is executed, and whether the corresponding parsed result is returned is checked.
- A query statement containing a production batch identifier is entered, and whether retrieval results automatically associate the document collection for the corresponding batch is confirmed.
- The vector database monitoring panel is viewed, and whether GPU resource usage shows corresponding fluctuations after retrieval requests are triggered is confirmed.
- A query with reference language such as "the last AOI report" is tested, and whether the retrieval results return the correct batch-corresponding documents is confirmed.
The values provided are common starting points and should be measured against the reader's own samples.