What Data for This Category Looks Like
Data for this category comes from Gerber files, drill files, and BOM provided by clients, plus internal factory CAM files, equipment operating parameters, bath solution test records, AOI inspection reports, electrical test reports, and laboratory slice reports. Data update frequency is per production batch, per shift, or continuous collection. It covers multiple types of documents including engineering production instructions, drilling lamination parameters, and chemical analysis results. Fields include specific values of process parameters, coordinates and defect types of test results, with units such as rotational speed, concentration, and length, following industrial standard units.
Constraints Imposed by These Characteristics on Vector Models and Indexing
Multi-source, high-frequency updated data requires indexes to support incremental updates, to avoid time loss caused by full reconstruction. Differences in structure across document types require vector chunking to adapt to both short process parameter fields and long paragraphs of inspection reports, preventing critical information from being split apart. Production ledger summary needs to associate process data across multiple links, which places high requirements on vector recall accuracy, to avoid incorrect matching of unrelated batch data. Additionally, the data contains a large number of precise numerical values and parameter differences, so embedding models must adapt to subtle feature differentiation in industrial scenarios.
How to Set the Configuration
| Configuration Item | Recommended Setting Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | PCB process documents include short parameter records and long inspection reports. This range can cover complete process node information |
| Overlap Length | 100–200 characters | Prevents cross-chunk process logic from being split, ensuring continuity of vector associations |
maxToken | 8192 token | Adapts to the context limit of mainstream embedding models supported by FastGPT v4.17.0, covering the summary requirements of complete production ledgers |
| Similarity Threshold | 0.72–0.85 | Requires differentiation of subtle differences between similar process parameters, avoiding incorrect association of production data from different batches |
| Single Search Limit | Top 10 results | Production ledger summary needs to cover core process links. Excessive results will add redundant information to report generation |
Three Common Misconfigurations
- Phenomenon: A model mismatch error is returned when calling the knowledge base interface to generate an index, or vector results do not match expectations after index generation. Cause: The embedding model version is not fixed, or a general embedding model with poor adaptability to PCB process data is used.
- Phenomenon: Index query response times out when creating a QA split file collection via OpenAPI. Cause: Chunk length and search limit parameters are not adjusted, resulting in excessive index data volume per batch and excessive load on the vector recall link.
- Phenomenon: Duplicate process parameter entries appear in the generated production ledger report. Cause: The similarity threshold is set too low, causing similar parameters belonging to different production batches to be incorrectly associated and summarized.
How to Verify Correct Configuration
- Upload a single drilling parameter document, check that the chunked vector fragments fully include core fields such as rotational speed and feed rate.
- Call the knowledge base search interface, verify that the number of returned results matches the single search limit setting.
- Compare vector similarity generated by different embedding models, confirm that the selected model can differentiate subtle differences between similar process parameters.
- Run an incremental index update task for consecutive production batches, check that the update time meets expectations.
The values provided here are common starting points and should be measured against the reader's own samples.