What the data for this category looks like
Typically, data for this category comes from manufacturing enterprise ERP orders and inventory, BOM lists, process routes, equipment capacity ledgers, staff shift schedules, work-in-progress inventory, and supplier arrival records. Daily processed documents include sales orders, master production schedules, material requirements plans, weekly and daily production schedules, work orders, material shortage lists, and delivery change notices. The document structure mixes text paragraphs and structured tables. Fields include order number, material code, capacity value, shift duration, work-in-progress quantity, with corresponding units such as pieces, hours, sets, and equipment sets, etc. Data update frequencies include daily, weekly, or real-time triggered by orders or abnormal situations.
What constraints these characteristics impose on the document parsing and chunking step
The mixed structure of multiple document types requires the parsing process to preserve the integrity of both text and tables, avoiding splitting key business fields. The high-frequency update feature requires the parsing workflow to have matching response speeds, avoiding delays that impact scheduling adjustments. Specific business fields and units must retain their associated relationships during chunking, preventing the separation of core logic such as material requirements planning and capacity allocation. Temporary documents in abnormal scenarios need to quickly identify key change information, ensuring that parsing results can be directly used for ledger summary.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
customPdfParse | Enabled | Adapts to the mixed table and text structure of manufacturing documents, supports MinerU enhanced parsing capabilities |
chunkSize | 800–1200 characters | Matches the business block length of most production documents, avoids splitting key row content of work orders and production schedules |
chunkSplitter | Split by paragraph + table | Preserves logical associations of production documents, avoids separating material requirements plan entries and their corresponding values |
trainingType | Table vector separation mode | Enables multi-vector support for structured tables in production documents, retains independent row and column semantics |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Adapts to parsing time for batch production documents, avoids parsing timeouts in high-frequency update scenarios |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Accommodates batches of order and production schedule collections, matches the upper limit of daily enterprise document scale |
Three common mistakes
- Phenomenon: Parsing tasks return timeout or failure after enabling
customPdfParse. Cause: The local MinerU API port is not configured correctly, or network connectivity between FastGPT and MinerU is restricted. - Phenomenon: PDF parsing effect shows no obvious change after selecting
customPdfParsein the knowledge base settings. Cause: The MinerU enhanced parsing function is not enabled in the FastGPT global settings, or the plugin has not been deployed. - Phenomenon: Parsed production schedules and BOM lists cannot accurately recall capacity and material values within tables. Cause:
trainingTypeis not configured as the table vector separation mode, resulting in table content being merged into ordinary text chunks.
How to confirm the configuration is correct
- Log in to the FastGPT v4.17.0 backend, navigate to the global settings page, confirm that the MinerU enhanced parsing function is enabled.
- Navigate to the knowledge base settings page, confirm that the configuration options for
customPdfParseandtrainingTypehave been adjusted as required. - Upload a local weekly and daily production schedule PDF, initiate parsing and review the results, confirm that table content is split into independent vector chunks.
- Call the FastGPT file parsing API, pass in a test material requirements plan document, verify that the returned chunk results fall within the
chunkSizeconfiguration range.
The values provided are common starting points and should be measured against the reader's own samples.