What Data for This Category Looks Like
The data for this category comes from engineering, production, quality control, and laboratory teams at PCB manufacturing factories. It covers engineering production instructions, Gerber parsing records, drilling and lamination parameters, bath solution test results, AOI reports, electrical test reports, and sectioning reports. Data sources include Gerber files, drill tapes, and BOMs provided by customers, as well as CAM-generated files, equipment operating parameters, and on-site inspection data. Updates occur per production batch, per shift, or via continuous collection. Individual documents contain both structured and unstructured fields such as parameter values, test results, and process records.
What Constraints Do These Characteristics Impose on the Vector Models and Indexing Stage
The multi-source, heterogeneous data structure of this category requires vector indexes to support mixed-format parsing for Gerber files, structured parameter tables, unstructured test reports, and other formats, to avoid feature loss caused by single parsing rules. The high update frequency requires indexes to support incremental synchronization logic, so that new test and parameter data from production batches can be covered without full reconstruction each time. The document structure that combines precise parameters and qualitative descriptions requires targeted configuration of vector encoding rules, to distinguish the feature weights between numerical process parameters and qualitative abnormal descriptions. Additionally, the associated attributes of process flow require indexes to retain contextual connections between documents, to enable quick tracing of the upstream process source of abnormalities during audits.
How to Set Configurations
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the information density of single blocks in PCB documents such as process parameters and test reports, avoiding information breakage within blocks |
| Overlap Length | 100–200 characters | Retains process context between adjacent chunks, preventing loss of cross-block parameter associations |
| Similarity Threshold | 0.75–0.85, calibrated via actual testing | Distinguishes the feature matching degree between normal process parameters and abnormal test results |
| Single Search Limit | Top 10 results | Covers the number of associated process documents required for single-batch audits |
maxToken | 4096 token | Adapts to the total length of PCB documents, avoiding truncation of critical Gerber parsing records |
defaultToken | 150 token | Handles vector encoding of short single parameter records, ensuring feature integrity for small-volume data |
Three Common Mistakes
- Phenomenon: After indexing hundreds of thousands of PCB data entries in the knowledge base, search response times exceed 30 seconds. Cause: Incremental index rules are not configured, and full vector matching is triggered for each search, leading to excessive computational load.
- Phenomenon: A large number of irrelevant bath solution test results appear in search results, making it impossible to locate drilling parameter abnormalities. Cause: Differentiated vector encoding weights are not configured for numerical parameters and text reports, leading to feature confusion.
- Phenomenon: After importing Gerber files, drilling coordinate fields are missing from index results. Cause: Chunk length is set too small, splitting complete coordinate parameters into multiple chunks, leading to truncation of critical information.
How to Verify Proper Configuration
- Import a single typical PCB process document, check that the chunked text retains complete parameter and test descriptions, with no critical information breakage.
- Initiate a search for abnormal process keywords, verify the number of returned results and the matching degree of associated processes, and adjust the similarity threshold to a range that meets audit requirements.
- Import newly generated production batch data, verify that the index only synchronizes new documents, and completes updates without requiring full reconstruction.
- Check the vector storage usage space, confirm that the
maxTokenand chunk length configurations do not cause individual vector data to become too large, exceeding storage limits.
The values provided are common starting points and should be measured against the reader's own samples.