What the data for this category looks like
Core data for the survey and design industry comes from project survey reports, professional design descriptions, administrative approval documents, change notices, and similar materials. Data updates follow project progress. Full documents are uploaded in bulk when a project starts. Revised content is pushed incrementally during mid-project changes. Most documents have a multi-level nested structure. This includes main text chapters, embedded tables, calculation formulas, image attachments exported from CAD, and supporting doc-format description files. Fields include project numbers, survey point coordinates, standard specification numbers, load parameters, and more. Units involve professional measurement standards such as meters, megapascals, and square meters.
What constraints these characteristics impose on the "document parsing and chunking" step
Multi-level nested document structures require retaining hierarchical associations during parsing. Without this, specific chapter questions in consulting work cannot be matched accurately. Single long documents can reach dozens of pages. Fixed-length chunking easily splits the context of professional parameters. This reduces retrieval accuracy. Non-standard format content such as embedded CAD images and doc attachments requires special parsing logic to extract text information. Large files uploaded in bulk take longer to parse. Adaptations for timeout and size limits are needed. Professional fields have strong associations between units and parameters. Key information combinations cannot be split arbitrarily during chunking.
How to configure the settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 1000–1500 characters | Matches the average length of single chapters in survey documents, avoiding overloaded or overly fragmented single chunks |
chunkSplitter | Split by title hierarchy | Aligns with the multi-level table of contents structure of documents, retaining contextual associations between chapters |
customPdfParse | Enabled | Supports parsing embedded CAD images, doc attachments, and nested tables, adapting to the special format requirements of survey and design documents |
UPLOAD_FILE_MAX_SIZE | 2000 MB | Adapts to bulk uploads of full project reports, preventing upload failures for large files |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Accommodates the time required for long document parsing, preventing interruptions to the parsing process due to timeouts |
chunkSettingMode | Manual configuration | Adjusts parameters for the special structure of professional documents, meeting the requirements of survey and design professional content |
Three common mistakes
- Symptom: Content in multi-level directories cannot be fully parsed or retrieved. Cause: Chunking rules are not configured to split by title hierarchy. Only splitting by fixed length causes loss of hierarchical information.
- Symptom: The file parsing API call shows success but no returned results. Cause: The
PARSE_FILE_TIMEOUT_SECONDSparameter is not set correctly. Long document parsing takes longer than the threshold, resulting in incomplete returned results. - Symptom: A marker error reports a split exception. Cause: The
chunkSizevalue is outside a reasonable range, orcustomPdfParseis not enabled, preventing normal parsing of embedded CAD images and triggering a split error.
How to confirm the configuration is correct
- Upload a survey PDF document that includes multi-level directories, and check if the parsed chunk results retain chapter hierarchical information.
- Call the file parsing API, and verify that the returned results include text content from CAD images and embedded doc attachments in the document.
- Upload a survey project report larger than 1000 MB, and confirm that the upload and parsing processes do not trigger size limit or timeout error messages.
- In FastGPT v4.17.0, adjust the
chunkSizeparameter and compare chunking results, ensuring that professional parameters are not arbitrarily split or fragmented.
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 2026-09-28.