What This Category of Data Looks Like
Daily documents for this category include packaging process cards, bonding and molding parameter tables, wafer incoming material drawings, test program version files, yield reports, failure analysis reports, and more. Data sources cover wafer test maps, MES systems, bonding and molding equipment, ATE testers, X-ray/SAM test results, and customer specification requirements. Data updates occur per production batch, per tested chip, and daily. Most documents include fields such as batch numbers, parameter values, test bin codes, and yield values. Some drawing documents include coordinate and dimension units.
Constraints on Document Parsing and Chunking for This Category
This category of documents contains mixed content: structured parameter tables, drawing files, and text reports. Accurate matching of production batches to corresponding test data is required, so the chunking process must avoid splitting associated fields. Frequently updated batch data requires parsing speed to meet real-time call needs. Documents with large per-batch data volumes must strictly control chunk boundaries to prevent cross-batch information mixing. Visual documents such as wafer incoming material drawings must support parsing and association of non-text content. Version files such as test program versions must retain the binding relationship between version numbers and corresponding parameters, avoiding loss of key associated information after chunking.
How to Configure
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the field density of text documents such as packaging process cards and yield reports, avoiding splitting the association between batch numbers and corresponding parameters |
chunkSplitter | By paragraph + semantic boundary | Adapts to documents mixing text parameters and drawing descriptions, avoiding truncation in the middle of key parameter lines |
customPdfParse | Enabled | Supports parsing visual content embedded in PDFs such as wafer incoming material drawings, extracting text annotation information associated with drawings |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Adapts to parsing time for merged uploads of multiple documents per batch, preventing parsing timeout for large-volume yield reports |
chunkSettingMode | Manual configuration | Adjusts chunking rules for different document types (parameter tables, reports, drawings) to meet the unique field association requirements of this category |
Three Common Configuration Mistakes
- Phenomenon: After uploading multiple packaging and testing documents, some documents cannot be retrieved, and search results only cover a single document. Cause: The
customPdfParseconfiguration is not enabled, resulting in PDF documents embedded with wafer test maps not being fully parsed, only extracting plain text fragments. - Phenomenon: Cross-production batch parameter confusion occurs during question answering, with two different batches of bonding parameters included in the same answer. Cause: The
chunkSizesetting exceeds 1500 characters, causing parameter lines of adjacent batches to be merged into the same chunk, making it impossible to distinguish batch attribution during retrieval. - Phenomenon: After associating with the MES database, the generated query statement cannot match the unique bin code field of this category, returning empty query results. Cause: The binding relationship between fields and their corresponding business meanings is not retained during chunking, resulting in loss of bin code identification information during metadata extraction.
How to Verify Correct Configuration
- Upload a single packaging process card PDF, enter the knowledge base parsing details page, view the chunk list, and confirm that each chunk contains complete association information between a single parameter line and the corresponding batch number.
- Upload a PDF document containing a wafer test map, check whether the parsing result includes the text annotation content of the drawing, and verify that the
customPdfParseconfiguration is effective. - Upload a single 150MB yield report, wait for the parsing task to complete, and confirm that the task status is successful with no timeout-related errors.
- Set
chunkSizeto 800 characters, retrieve single-batch yield data, and confirm that the retrieved results only include parameter content of that batch, with no cross-batch parameter mixing.
The values provided are common starting points and should be evaluated against the reader’s own samples.