What data for this category looks like
Data sources for this category include the unified budget management system, project application materials, fund payment data, business system ledgers, and satisfaction surveys. The update frequency is annual. Documents processed daily include performance target forms, self-assessment forms, performance evaluation reports, supporting material lists, and rectification reports. Most document structures consist of tables nested with free text. Some supporting materials are in scanned document format. Fields include performance indicators, fund association identifiers, rectification items, and corresponding associations need to be identified using document context.
What constraints do these characteristics impose on the "document parsing and chunking" link
The annual centralized update feature requires batch processing of multiple documents, which creates requirements for the concurrency capability of parsing tools. Documents contain nested tables and structured fields. When chunking, the context of table rows and associated text must be retained to avoid separating performance indicators from their corresponding supporting explanations. There is a one-to-one correspondence between supporting materials and evaluation reports. Chunking must retain cross-document association identifiers. Some documents are in scanned format, so OCR parsing support is required to extract valid content. Otherwise, a large amount of structured information will be lost.
How to configure
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Performance evaluation reports contain long paragraphs and nested tables. This range can retain the associated context of indicators and supporting materials |
chunkSplitter | Split by paragraph + table boundaries | Documents contain structured tables and free text. Splitting by this rule avoids separating performance indicators from their corresponding supporting explanations |
customPdfParse | Enable OCR mode | Some supporting materials are scanned PDFs. Enabling this mode can extract scanned content |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Single performance evaluation reports have relatively long length. This duration covers the complete parsing process |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Fiscal performance documents often contain batch supporting materials. This upper limit covers the maximum file volume for a single batch upload |
chunkSettingMode | Custom mode | Need to adjust chunking rules for the structured characteristics of fiscal performance documents to adapt to different types of performance documents |
Three common mistakes
- After passing in a public document link, the parsing result is empty and only returns page metadata. Cause: The
chunkSplitteris not configured to adapt to the structured extraction rules for web content, and the default parsing rules cannot capture the main body of fiscal performance-related documents. - The chunk ID on the knowledge base page cannot be selected and copied. Cause: The interface does not open interactive permissions for this field, and no adaptive configuration is made for the ID display of structured documents.
- After custom chunking, documents are stored in the knowledge base, and duplicate blocks are automatically deleted, causing the index order to be disrupted. Cause: The knowledge base enables the document chunk deduplication logic by default, and this configuration is not turned off in advance to retain the complete order of custom chunking.
How to confirm the configuration is correct
- Upload a performance self-assessment form in scanned format, check if the parsing result includes the performance indicators and numerical content in the table.
- Adjust
chunkSizeto 1000 characters, upload a long document, check if each chunk after chunking retains the complete performance indicator paragraph. - Upload batch documents, check if the parsing time is within the range set by
PARSE_FILE_TIMEOUT_SECONDSwith no timeout errors. - After enabling custom chunking, upload a document, confirm that the number of blocks stored in the knowledge base matches the number after custom chunking.
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.