What the data for this category looks like
The marketing documents for this category mainly originate from packaging and printing promotional materials of financial, insurance, and wealth management institutions, including product brochures, contract packaging manuals, customer rights notifications, wealth management plan sample books, and similar materials. Document updates are triggered irregularly alongside business adjustments: new or revised documents are added when new products launch, compliance requirements are updated, or pricing is adjusted. Most documents use a mixed text and image layout; some include actual sample photos or scanned paper sample book content. Built-in fields include packaging material, grammage, finished product size, printing process, quoted unit price, and more, with corresponding units of g/㎡, mm, yuan/thousand units, individual units, and similar units.
What constraints do these characteristics impose on the document parsing and chunking link
The mixed text and image layout with scanned content requires the parsing process to support OCR recognition of text embedded in images, to avoid losing parameter descriptions from actual sample books. The characteristic that fixed fields have dedicated units requires that the binding relationship between parameters and units be retained during chunking, and parameters and their corresponding metric values must not be split. Documents are divided into semantic blocks such as materials, quotations, and cases based on business logic, and should not be split solely by character length. The irregular update of materials requires that parsing and chunking support incremental processing to adapt to quick synchronization of new or revised documents. Some documents include long tables of quotation details, and long tables must be prevented from being unnecessarily truncated, which would cause semantic breaks.
How to set the configurations
| Configuration Item | Recommended Value | Rationale for This Setting |
|---|---|---|
ocr_enabled | Enabled | Most packaging and printing documents include actual photos and scanned materials, requiring recognition of embedded parameter and descriptive text |
chunk_size | 800–1200 characters | Documents include long parameter tables and business descriptions; this range preserves semantic integrity and avoids splitting key business blocks |
chunk_overlap | 10–15% | Cross-block parameter associations must be retained, preventing key information such as materials and sizes from being truncated at chunk boundaries |
parse_scanned_pdf | Enabled | A large number of old packaging sample books use scanned PDF format, requiring extraction of embedded text content |
UPLOAD_FILE_MAX_SIZE | 200 MB | Supports complete upload of multi-page thick sample books and high-definition sample book images |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Parsing thick scanned documents or multi-page sample books requires a longer processing cycle |
The parameter values provided on this page are all common recommended starting points for determining configurations. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Phenomenon: Key parameter content such as packaging materials and quotations cannot be retrieved during knowledge base search, or the retrieved results do not match the search keywords. Cause: OCR recognition of scanned documents or mixed text and image content is not enabled during chunking, resulting in key parameters not being extracted into the chunked text.
- Phenomenon: Parsing tasks remain in the "parsing" state for a long time and eventually return a timeout error. Cause: The
PARSE_FILE_TIMEOUT_SECONDSparameter is not adjusted, and the default timeout period is insufficient to handle the parsing process of thick scanned sample books. - Phenomenon: Chunked results split complete quotation tables into multiple unrelated fragments, making it impossible to obtain complete quotation information during search. Cause: Chunking is performed only based on fixed character length, not split based on business semantic blocks, and a reasonable chunk overlap rate to retain cross-block associations is not configured.
How to confirm that configurations are properly set
- Upload a packaging sample book PDF that includes scanned content, check the parsed text content, and confirm that the parameter text in the actual photos has been correctly extracted.
- Perform a knowledge base search test, enter keywords for dedicated fields such as material grammage and printing process, and confirm that chunked results containing the corresponding content can be retrieved.
- View the chunk detail page, confirm that long tables or parameter blocks are not unnecessarily truncated, and that adjacent chunks have reasonable overlapping content.
- Upload a multi-page packaging sample book, confirm that the upload and parsing processes complete normally without error prompts.
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.