What Data for This Category Looks Like
Data for this category comes from wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification documents. Daily documents include packaging process cards, bonding and molding parameter sheets, wafer incoming material maps, test program version files, yield reports, and failure analysis reports. Data is updated per production batch, after each individual chip test, and daily. Documents contain structured and semi-structured fields including process parameters, batch identifiers, electrical test bin code distributions, wafer coordinates, equipment operating parameters, and more.
Constraints Imposed on Document Parsing and Chunking
Wafer incoming material map coordinate parameters must precisely match physical production batches; parsing must retain spatial coordinate association information. Yield report bin code distributions must be chunked by test batch to avoid cross-batch data confusion. Daily updated equipment operation logs must be split by timestamp to maintain contextual association. Format differences across multi-source data (MES, ATE test machines) require tailored parsing rules to prevent field misalignment. Single-batch document size increases with production scale; single chunk content length must be limited to avoid context overflow.
Configuration Specifications
These configuration rules are based on FastGPT v4.17.0. The following table details the settings:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the field density of packaging process cards and yield reports, prevents single chunks from containing cross-batch data |
chunkSplitter | Chunk by paragraph + semantic meaning | Adapts to structured paragraph splitting for bonding parameter sheets and failure analysis reports, retains parameter association logic |
customPdfParse | Enabled | Adapts to vector format parsing of wafer incoming material maps, retains accuracy of coordinate fields |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Adapts to parsing duration of large single-batch test program version files, avoids timeout interruptions |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Accommodates batch upload requirements for multiple ATE test logs and wafer test maps |
chunkSettingMode | Manual configuration | Adjusts chunking rules for different document types (e.g., process cards, test reports) to match business data characteristics |
Three Common Configuration Mistakes
- Parsing result is empty after uploading a wafer incoming material map. Cause: The
customPdfParseconfiguration is not enabled. Default PDF parsing cannot retain vector coordinate fields, resulting in failed structured data extraction. - Parameter error is returned when calling a workflow API to upload images. Cause: The binary stream of the image file or correct form field name is not passed per interface requirements, leading to parameter verification failure.
- Chunk results include cross-batch yield data. Cause:
chunkSizeis set too large, merging yield reports from multiple production batches into a single chunk, causing confusing data association during subsequent retrieval.
How to Confirm Proper Configuration
- A single packaging process card is uploaded. The parsed text is checked for core fields including process number, parameter values, and batch number to confirm
customPdfParseis active. - The file parsing API is called, and a test wafer incoming material map is passed. The returned result is checked for wafer coordinate parameters to confirm the semantic chunking logic of
chunkSplitteris correct. - An ATE test log package within 1000 MB is uploaded. The parsing task is confirmed to complete within the duration set by
PARSE_FILE_TIMEOUT_SECONDSwith no timeout errors. chunkSizeis manually adjusted to 600 characters and 1200 characters. Chunk results are compared to confirm single chunks do not span multiple independent documents.
The values provided are common starting points and should be measured against the reader's own samples.