What data for this category looks like
The core entities for this category include raw material warehouses, line-side warehouses, finished product warehouses, and logistics material control departments. Daily generated documents include receiving slips, putaway slips, material picking and return slips, warehouse transfer slips, inventory checklists, lot ledgers, first-in-first-out and expiration date records, and more. Data sources cover ERP/WMS systems, barcode or RFID equipment, procurement and work order information, warehouse location data, AGV or conveyor systems, weighing data, and others. Data update frequencies range from real-time, daily summaries to monthly inventory cycles. Most documents are structured forms with fixed fields including lot numbers, warehouse location codes, expiration dates, material quantities, weights, and more. Some exception reports contain unstructured on-site descriptions and rectification suggestions.
What constraints these characteristics impose on document parsing and chunking
Warehousing and logistics documents contain a large number of associated fields, such as lot numbers bound to expiration dates, and warehouse locations corresponding to material quantities. Chunking must avoid splitting these associated pieces of information, otherwise the accuracy of subsequent process document and exception report drafting will be affected. Real-time updated documents require fast parsing to avoid impacting the timeliness of logistics distribution. Batch documents for monthly inventory checks require the system to handle large file volumes to prevent parsing timeouts. Some documents may be exported from different systems with minor format differences, so parsing rules need a certain level of adaptability. Unstructured content in exception reports must retain contextual logic to ensure complete information during drafting.
How to set configurations
These configurations are adapted for FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the field length of warehousing documents, avoids splitting associated fields such as lot numbers and warehouse location codes, and ensures complete information within a single chunk |
chunkSplitter | Chunk by paragraph + semantic meaning | Retains contextual association between on-site descriptions and rectification suggestions in exception reports, and avoids breaking logical coherence of process steps |
customPdfParse | Enable OCR-assisted parsing | Handles some scanned inventory checklists or scanned paper documents, ensuring printed or handwritten fields can be correctly identified |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Meets the upload requirements for batch monthly inventory documents, preventing upload failures due to oversized files |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Addresses parsing time for large batch ledgers, preventing parsing interruptions due to timeouts |
chunkSettingMode | Manual configuration | Adjusts chunking rules to match business requirements for the fixed field structure of warehousing documents |
Three common mistakes
- After deployment, prompts indicate insufficient video memory, or the CUDA 12.8 environment cannot be detected. The cause is failure to mount GPU devices in the Docker startup command, or failure to specify a CUDA runtime version matching the graphics card driver.
- Parsed lot ledger fields are missing, or warehouse location codes are split across different chunks. The cause is an overly small
chunkSizesetting, which forces splitting of associated fields and destroys information integrity. - After uploading an exception report with images, the image links in the parsed Markdown document lack a domain name prefix. The cause is failure to configure custom file parsing rules, and failure to retain domain information from the original upload path of images.
How to confirm configurations are set correctly
- Upload a standard receiving slip PDF, check that parsed chunks retain complete association between lot numbers and warehouse location codes, with no field splitting occurring.
- Test batch uploads of monthly inventory checklists, confirm that the parsing process does not experience timeout interruptions, and that elapsed time meets expectations.
- Upload an exception report with images, check that image links in the parsed Markdown document retain the original domain name prefix.
- Review system operation logs, confirm that GPU resources are being called normally, with no error messages for video memory overflow or device detection failure.
The values provided are common starting points, and should be evaluated against the user's own samples.