What data for this category looks like
Core data for this category comes from product help centers, customer organizational structures, training recording transcripts, and ticket systems. Update cycles adjust based on individual customer version launches or single training cycles. Daily processed documents include standardized user manuals, structured training slides, training attendance sheets, and frequently asked question (FAQ) answer lists. Most document structures include three module types: operation steps, permission descriptions, and common Q&A. Fields mostly correspond to customer roles, operation nodes, and Q&A scenarios. No unified fixed units exist. Values adjust flexibly based on customer business needs.
What constraints do these characteristics impose on document parsing and chunking
The data characteristics of this category impose multiple constraints on parsing and chunking. Short text fragments from ticket system FAQs require smaller chunk sizes to ensure retrieval accuracy. Long documents generated from training recording transcripts contain many cross-chapter operation step links. Chunking must retain contextual connections between adjacent paragraphs. Document structures vary significantly across different customers. Custom chunking rules must be supported to adapt to non-standard formats. Frequently updated documents require automatically triggered parsing configurations to avoid content lag. The core goal is to help customers master product usage methods. Chunked content must make it easy to retrieve operation steps and role permission descriptions. Failure to do this directly impacts training implementation effectiveness.
How to set configurations
This configuration guide is written for FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the chapter and paragraph length of user manuals and training slides, retains complete context for operation steps and permission descriptions |
chunkSplitter | Split by heading hierarchy | Matches the heading structure of chapters and subsections in documents, avoids splitting cross-module content |
customPdfParse | Enabled | Adapts to common image and table mixed formats in training slides and user manuals, fully extracts text and embedded resources |
UPLOAD_FILE_MAX_SIZE | 500 MB | Compatible with large files generated from single-batch training recording transcripts, prevents upload failures |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Adapts to time requirements for long document parsing, prevents parsing interruptions due to timeout |
chunkSettingMode | Custom mode | Adapts to differences in document structures across different customers, supports adjusting chunking rules as needed |
Three common configuration mistakes
- Observation: No embedded images appear in parsing results after uploading a PDF. Cause: The
customPdfParseconfiguration item is not enabled. The default parsing only extracts plain text and ignores image and table content. - Observation: Retrieval results based on user manuals and training slides have low relevance. Cause: The
chunkSizesetting is too large or too small, causing chunked content to fail to match operation step fragments that users actually search for. - Observation: Long document parsing tasks terminate early. Cause: The
PARSE_FILE_TIMEOUT_SECONDSsetting value is lower than actual parsing time, triggering timeout interruptions.
How to confirm configurations are correctly set
- Upload a test user manual PDF, and verify whether the parsing result includes images and table content from the original document.
- Submit a long document parsing task, wait for the duration set in
PARSE_FILE_TIMEOUT_SECONDS, then check whether the task status is completed. - Adjust
chunkSettingModeto custom mode, attempt modifying chunking parameters, and verify whether the configuration takes effect. - Upload a single test document, and verify whether the number of generated chunks matches the number of document chapters.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.