What the Data for This Category Looks Like
This category’s data is sourced from engineering, production, quality control, and laboratory teams at printed circuit board factories. Daily documents include engineering work instructions, Gerber analysis records, drilling and lamination parameters, chemical analysis reports, AOI inspection reports, electrical test reports, and slice analysis reports. Data sources include customer-provided Gerber files, drill tapes, BOM lists, CAM files, equipment operating parameters, bath test results, and various types of inspection data. The update frequency is per production batch or per shift. Some inspection data is collected continuously.
Constraints Imposed by These Characteristics on Vector Models and Indexes
The multi-source, heterogeneous nature of this category’s data requires vector indexes to support text converted from original documents of different formats. This includes process parameters, inspection reports, customer-provided engineering files, and other document types. The high-frequency update rhythm requires indexes to support incremental updates, avoiding full reindexing that consumes excessive server resources. The wide range of document lengths — from short drilling parameter entries to hundreds-of-line slice analysis reports — requires flexible chunking rules. The bid scenario requires precise matching of customer-specified process requirements and quality standards. Retrieval must strictly link to corresponding business fields to avoid irrelevant content interfering with matching results.
How to Configure Parameters
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the span of PCB documents from short parameter entries to long inspection reports, avoiding overly fragmented chunks that disrupt business logic or overly long chunks that exceed token limits |
| Overlap Length | 50–80 characters | Retains process-related information across chunks, such as connecting content between drilling parameters and subsequent lamination requirements |
| Similarity Threshold | 0.75–0.85 | Meets the precision requirements for process parameters in bid scenarios, filtering low-relevance non-target document content |
| Single Search Limit | Top 8 results | Controls the scale of retrieval results, adapting to the need for multi-parameter verification during bid document preparation |
defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
Three Common Misconfigurations
- Phenomenon: After uploading a large number of PCB process documents and customer materials, the server crashes. After restarting, the knowledge base status remains "Not Ready", and no automatic indexing is triggered. Cause: Incremental indexing trigger rules are not configured. Full indexing tasks use too many server resources during high-concurrency uploads. The breakpoint resume indexing logic is not restored after restarting.
- Phenomenon: After upgrading FastGPT v4.17.0, the original indexing knowledge base configured for PCB parameters cannot return retrieval results. Cause: The default chunking rules for the vector model changed after the version update, and the chunk length and overlap parameters for PCB documents were not readapted.
- Phenomenon: Retrieval results include a large amount of irrelevant bath test data, and cannot accurately match the drilling parameters required for the bid. Cause: The similarity threshold is not set, or the threshold is set too low, and low-relevance non-process document content is not filtered.
How to Confirm the Configuration Is Correct
- Upload a single typical PCB engineering work instruction document, and check whether the indexing progress bar completes normally with no timeout errors.
- Enter keywords for a simulated bid requirement, such as "0.2mm drilling parameters", and verify whether the retrieval results include relevant content from the corresponding process documents.
- After restarting the server, check whether the knowledge base status automatically switches to the ready state, with no manual indexing operation required.
- View the vector indexing operation logs, and confirm that token consumption does not exceed the preset
maxTokenlimit, with no resource overflow errors.
The values listed above are common starting points, and should be evaluated against the user’s own sample data.