What this category's data looks like
Typical entities include cotton spinning, chemical fiber spinning, woven and knitting enterprises. Daily documents include cotton blending or batching lists, process setup sheets, equipment operation records, yarn inspection records, grey fabric inspection records, and defect records. Data sources cover raw fiber inspection data, ERP systems, spinning and loom operation data, evenness tester data, fabric inspection machine collected data, and order specification information. Updates sync with each batch of raw materials, each shift's production, and each piece of fabric output. Document fields include fiber fineness, yarn count, twist, weaving structure parameters, defect types, and more. Units include tex, twists per meter, piece count, and others.
Constraints on document parsing and chunking
Documents in this category have high update frequency, diverse types, and highly specialized fields. Chunking must retain complete business context. Cross-batch cotton blending lists and cross-shift equipment records must not be mixed in the same chunk. Different documents have large structural differences: cotton blending lists use structured table formats, while defect records use unstructured text formats. Multiple splitting rules must be supported. Large PDF process sheets and multi-page equipment records consume significant parsing resources. Single chunk size must be controlled to balance parsing efficiency and context integrity.
Configuration settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | A single chunk can contain complete process parameters or defect records, avoid context breaks, and meet the accuracy requirements of business queries |
chunkSplitter | Split by paragraph + fixed character | Balances chunking needs for structured table documents and unstructured text documents, avoids splitting the same business unit into multiple chunks |
customPdfParse | Enabled | Adapts to scanned or embedded font PDF process sheets and defect records, extracts complete embedded text layers, and resolves issues where some PDFs cannot be recognized |
UPLOAD_FILE_MAX_SIZE | 500 MB | Supports batch upload of multiple inspection records and multi-page process sheets matching a single batch of raw materials, adapts to daily document scales of enterprises |
PARSE_FILE_TIMEOUT_SECONDS | 120–180 seconds | Prevents parsing timeouts for large PDF documents, adapts to parsing durations for complex process sheets |
chunkSettingMode | Manual mode | Allows adjusting chunking rules for different document types such as cotton blending lists and equipment records, meeting the needs of specific business scenarios |
Three common configuration errors
- Symptom: An error is returned when calling the document parsing tool after uploading a local file. Cause:
UPLOAD_FILE_MAX_SIZEis not configured to adapt to large process sheets, or the file path is not correctly bound to the workflow input node. - Symptom: Some PDF process sheets fail to recognize valid content. Cause:
customPdfParseis not enabled; the default parsing only extracts visible surface text, and cannot handle scanned or encrypted PDF documents. - Symptom: Chunk results include irrelevant business content across batches. Cause:
chunkSplitterdoes not split by document batch markers, mixing cotton blending lists from different production batches into the same chunk, leading to context confusion during queries.
How to verify correct configuration
- A single 100 MB PDF process sheet is uploaded, and parsing progress is checked to confirm completion within
PARSE_FILE_TIMEOUT_SECONDS. - A standard cotton blending list document is imported, and chunk results are checked to confirm inclusion of complete fiber indicators and batching dosage fields, with no cross-chunk breaks.
- After enabling
customPdfParse, a scanned defect record PDF is uploaded, and parsed content is checked to confirm extraction of defect types and corresponding grey fabric numbers. - The file upload node and document parsing tool are bound in a FastGPT v4.17.0 workflow, and a local file is uploaded to test generation of correct chunk results.
The values provided are common starting points and should be measured against the reader's own samples.