What This Category of Data Includes
Daily documents generated by engineering, production, quality control, and laboratory departments in printed circuit board factories, including production instructions, Gerber analysis records, drilling and lamination parameters, chemical solution analysis, AOI reports, electrical test reports, and slice inspection reports. Data sources include customer-provided Gerber files, drill tapes, BOMs, CAM system data, production equipment parameters, bath solution test results, AOI and electrical test equipment data, and laboratory test records. Update frequency is per production batch, per shift, or continuous collection. Each document contains specific process parameters and test values, with structure varying significantly by document type.
Constraints Imposed on Vector Models and Indexing
Multi-source heterogeneous data sources and format differences require indexing systems to support semantic association across document types, preventing loss of key parameters due to format incompatibility. High-frequency updated production data requires adaptation to incremental indexing mechanisms, ensuring newly added drilling and lamination parameters are included in the retrieval scope in a timely manner. Specific values and continuous process steps contained in single documents require vector models to capture logical associations between parameters, preventing cuts that break process context. Business attributes with severe error consequences require index recall accuracy to meet strict process verification requirements, preventing irrelevant data recall that leads to incorrect judgments.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the semantic length of PCB process documents, matches the default setting of defaultToken, avoids losing parameter associations due to overly long chunks |
defaultToken | 1024 token | Carries the semantic content of a single chunk, covers most single-segment information in process documents |
| Overlap Length | 150–200 characters | Retains parameter associations between adjacent chunks, avoids cutting continuous process steps such as drilling and lamination |
| Similarity Threshold | 0.75–0.85, calibrated based on actual testing | Distinguishes subtle differences in process parameters, avoids recalling irrelevant bath solution analysis or quality inspection reports |
| Single Search Limit | Top 8 entries | Covers multi-link associated documents involved in PCB process queries, avoids recalling excessive redundant data |
maxToken | 4096 token | Carries the context and recall results of a single round of Q&A, avoids truncating key Gerber analysis record content |
Three Common Misconfigurations
- A 503 error occurs, prompting that text-embedding-v3 has no available channels under the default group. Cause: The channel group for the corresponding embedding model is not configured, or the call quota for the group is exhausted.
- Index parameters cannot be adjusted directly from the retrieval result page; configuration changes must be made on the knowledge base content page. Cause: The retrieval debugging entry is confused with the knowledge base configuration entry. In FastGPT v4.17.0, retrieval debugging only supports previewing recall results.
- Recall results contain a large number of irrelevant AOI reports. Cause: The similarity threshold is set too low, and the semantic difference between process procedures and quality inspection reports is not distinguished, resulting in insufficient recall accuracy.
How to Verify Proper Configuration
- Upload a single drilling parameter document, check whether continuous process steps are retained after chunking, with no obvious parameter cutting.
- Initiate a query containing specific process parameters, such as "What problems will abnormal lamination temperature cause", and check that the recall results include the corresponding lamination parameter documents.
- View the index update log, confirm that newly added data per production batch can be automatically indexed incrementally, with no update delay.
- Verify the calling status of the embedding model, with no 503-type errors, ensuring that the model channel configuration is normal.
The values provided are common starting points and should be measured against the reader's own samples.