What Data Looks Like for This Category
Data sources for this category include Gerber files, drill files, and BOM lists delivered by customers, as well as factory internal CAM system data, production equipment parameters, bath solution test records, AOI inspection reports, electrical testing data, and laboratory slice reports. Daily generated documents cover engineering production instructions, Gerber analysis records, drilling and lamination parameter tables, and chemical solution analysis logs. Data update frequency varies with production batches and shifts. Some equipment data is collected continuously. Each individual document includes clear parameter fields and units.
Constraints Imposed on Vector Models and Indexes
Diverse data sources and document formats require vector models to support embedding for both structured parameter text and professional technical descriptions. Frequently updated production data requires indexes to support incremental refresh and batch processing, to avoid excessive resource usage from full index reconstruction. Professional parameters and associated relationships within individual documents require vector embedding processes to preserve semantic connections between fields, preventing semantic loss from parameter splitting. The strong binding between production batches and data requires indexes to support retrieval grouped by production batch, improving the accuracy of cross-document association.
Configuration Settings
When using the configurations on this page, confirm that FastGPT v4.17.0 is in use. The defaultToken parameter of some embedding models must match the model's actual context window. Refer to the model's official documentation to confirm the correct value.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | PCB documents contain long parameter tables and professional technical descriptions. Excessive length damages the integrity of semantic connections, while insufficient length splits associated relationships between parameters |
| Overlap Length | 100–200 characters | Prevents continuous parameters or technical descriptions across chunks from being split, preserving contextual semantics |
maxToken | 8192 | Adapts to context window limits of most mainstream embedding models, ensuring single-chunk text can be fully embedded |
| Similarity Threshold | 0.72–0.85 | Balances retrieval accuracy and recall rate, adapting to parameter matching requirements for PCB professional documents |
| Maximum Results per Search | Top 8 entries | Reduces the number of candidate documents, avoiding unnecessary time consumption in subsequent reranking steps |
| Reranking | Enabled | Filters low-relevance retrieval results, improving the relevance of final matched professional documents |
Three Common Configuration Mistakes
- The interface displays "No available vector model channel". The cause is failure to correctly configure the API key and access address of the corresponding embedding model in the FastGPT system, or failure to bind the model to the default group.
- Vector embedding tasks time out. The cause is that chunk length is set too large, exceeding the model's
maxTokenlimit, leading to excessively long single-chunk embedding time. - The reranking step takes too long after retrieval. The cause is that the maximum results per search is set too high, and reranking is enabled, causing a large number of candidate documents to enter the reranking process.
How to Verify Correct Configuration
- Access the FastGPT model management page, check the access status of the embedding model, and confirm it shows normal online status.
- Upload a single typical PCB engineering production instruction document, trigger the vector embedding task, and confirm the task log contains no error messages.
- Enter a test query such as "Lamination Parameter Abnormality Troubleshooting", and confirm the retrieval results include the corresponding lamination parameter documents.
- Adjust the similarity threshold or chunk parameters, and verify that the relevance of retrieval results changes as expected.
Final Note
The values provided here are common starting points, and should be evaluated against samples relevant to the specific use case.