What data for this category looks like
Data for this category originates from engineering, production, quality control, and laboratory teams at printed circuit board factories. It covers engineering production instructions, Gerber parsing records, drilling and lamination parameters, bath test results, AOI reports, electrical test reports, and sectioning reports. Data sources include customer Gerber files, drill tapes, BOM lists, CAM files, equipment operating parameters, bath test data, and laboratory test results. Data is collected continuously per production batch and each shift. Individual documents contain fields such as parameter values, test times, batch numbers, and equipment identification numbers. Units include millimeters, ppm, ohms, and other specialized industrial units.
Constraints on the knowledge base retrieval and recall workflow
PCB manufacturing data is updated frequently and continuously. This requires the retrieval and recall workflow to support real-time or near-real-time indexing, to avoid review failures caused by delayed data. A mix of structured and semi-structured documents of multiple types means the retrieval workflow must support parameters and records in different formats. It cannot rely solely on generic text splitting logic. Individual documents have many specialized fields and unique units. Retrieval must preserve field associations, to avoid losing the binding between parameters and batch information after splitting. Production batch-level updates require the retrieval range to be filtered quickly by batch. It also requires limiting the number of documents retrieved per request, to prevent context overload that reduces review accuracy.
Configuration settings
| Configuration Item | Recommended Setting | Basis for This Setting |
|---|---|---|
similarity | 0.75–0.85 | Matches specialized parameters and document relevance for PCB manufacturing, filters low-relevance generic text |
limit | Top 8–12 entries | Controls the number of retrieved documents per batch review, prevents context overload that harms cross-verification |
searchMode | mixedRecall | Balances semantic matching for specialized parameters and keyword retrieval for structured files such as Gerber and BOM |
usingReRank | Enabled | Performs a secondary relevance sort on retrieved documents, corrects bias from single vector retrieval |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands search terms, covers variants of specialized terminology such as drilling compensation and lamination parameters |
collectionIds | Bind exclusive collections by production batch | Isolates documents from different production batches, ensures cross-review only uses valid data from the current batch |
Three common mistakes
- A 503 error occurs when uploading Gerber format files. This happens because the chunk threshold for knowledge base uploads is not adjusted for large binary files from PCB manufacturing.
- Low-relevance results are returned when searching for specialized parameters. This happens because
datasetSearchUsingExtensionQueryis not enabled, so variants of specialized terminology such as drilling compensation and lamination parameters are not covered. - Accurate answers cannot be retrieved for complex electrical connectivity issues. This happens because only flat entity retrieval is used, without preserving the binding relationship between Gerber graphics and drilling parameters.
How to confirm configurations are set correctly
- In the FastGPT v4.17.0 knowledge base management interface, check if currently bound collections are divided by production batch, to confirm the
collectionIdsconfiguration is correct. - Run a search for drilling parameters, verify that the similarity score of retrieved results falls within the 0.75–0.85 range.
- Upload a test Gerber file, confirm no 503 errors occur during upload, to verify chunk configuration works for large industrial files.
- Run a search requiring electrical connectivity checks, confirm retrieved results include associated documents for drilling parameters and lamination compensation, to verify
searchModeand reranking configurations are active.
The values provided are common starting points, and should be evaluated against the reader's own samples.