What this type of data looks like
Financial report data for this use case comes primarily from publicly disclosed annual, semi-annual, and quarterly reports, including attachments such as consolidated statements, segment operating data, and related party transaction details. Updates follow a fixed schedule: annual reports are released once per year, while semi-annual and quarterly reports are published per their respective cycles. Document structures typically include consolidated balance sheets, income statements, and cash flow statements, plus modules such as segment revenue breakdowns and related party transaction explanations. Fields cover items including net profit attributable to parent company owners, revenue proportions of each business segment, and more. Units are mostly marked in yuan, ten thousand yuan, or hundred million yuan; some cross-border enterprises include data converted from multiple currencies.
What constraints do these characteristics impose on document parsing and chunking?
The nested multi-module structure of financial reports requires that context links between consolidated statements and segment data be preserved during parsing, to avoid splitting core logic into isolated chunks. Large documents released on a fixed cycle require adaptation for long-text processing, while also identifying independent information units for different business segments. Fields marked with multiple currencies and units require the parsing process to capture unit information simultaneously, to prevent data confusion in subsequent analysis. Attachments containing detailed related party transactions, executive compensation, and other segmented content require accurate identification of module boundaries, to avoid invalid cross-module chunking that causes information disorder.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
max_chunk_size | 800–1200 characters | Financial reports include consolidated statements and segment data. This length can cover complete information for a single statement line or a single business segment, avoiding splitting critical data |
chunk_overlap | 150–200 characters | Preserves context links across chunks, such as segment revenue and corresponding segment note explanations, preventing loss of context |
parse_references | Enabled | Related party transactions and segment data often include reference annotations. Enabling this allows complete extraction of citation sources to support subsequent traceability analysis |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Large annual financial reports include multi-module attachments, resulting in long parsing times. This duration covers the full parsing process |
enable_pdf_marker | Enabled | Accurately identifies table structures in PDF files, avoiding cell misalignment in consolidated statements and improving chunking accuracy |
retain_original_units | Enabled | Financial reports use multiple unit markings. Retaining original units avoids data conversion errors |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require case-by-case analysis. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Symptom: The references field in parsed documents is empty or not extracted. Cause: The
parse_referencesconfiguration item is not enabled, or the special reference annotation format in financial reports is not adapted. - Symptom: Unable to find the corresponding setting item in the interface when trying to modify the default maximum chunk length. Cause: The
default_max_chunk_sizeparameter in the source code'sconfig/parse.jsfile is not adjusted, or the service is not recompiled and deployed. - Symptom: The pdf-marker function works normally locally, but no parsing entry appears when uploading a PDF to the FastGPT knowledge base. Cause: The pdf-marker integration is not enabled in FastGPT's environment configuration, or the container mount path is not correctly associated with the pdf-marker service port.
How to confirm configurations are set correctly
- Upload a small quarterly financial report document, check if the parsed text contains complete references field content, and verify that the configuration items are effective.
- Enter the FastGPT knowledge base configuration page, check if the chunk length setting matches the preset
max_chunk_sizeparameter, and confirm that parameter synchronization is normal. - Start the local pdf-marker service and associate it with the FastGPT container, upload a PDF document, check if the parsing entry displays normally, and verify that the integration configuration is correct.
- View the parsing logs, confirm that the timeout threshold corresponding to the
PARSE_FILE_TIMEOUT_SECONDSparameter does not trigger errors, and verify that the duration configuration matches the document volume.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.