Document Parsing and Chunking for Funding Source KYC

Funding source KYC data mainly comes from bank corporate/personal transaction statements, personal income certificates, asset holding documents

What this category of data looks like

Funding source KYC data mainly comes from bank corporate/personal transaction statements, personal income certificates, asset holding documents, annual tax returns, and similar materials. The update frequency depends on the verification scenario, with monthly or quarterly updates. Document formats include native PDF bank electronic statements, scanned paper certificates, and Excel-exported transaction detail tables. Core fields include transaction date, transaction amount, transaction counterparty, and transaction remarks. The primary unit is RMB yuan, and some cross-border scenarios include foreign currency fields.

What constraints do these characteristics impose on the "document parsing and chunking" step

The tabular transaction details and long-text structure of funding source documents require that chunking does not split associated fields across pages or chunks. Otherwise, KYC verification will fail to match transactions and their corresponding remarks. Scanned documents must first undergo OCR extraction to obtain valid text content. Large annual statement files may exceed the default upload threshold, so upload configuration adjustments are needed. Document formats vary widely across different sources, so compatibility with native PDFs, scanned documents, table-exported PDFs, and other forms is required to avoid field loss due to format incompatibility.

How to configure the settings

Configuration ItemRecommended ValueRationale
maxChunkSize800–1200 charactersFunding source documents mostly contain associated transaction fields. Overly long chunks will split transaction context, while overly short chunks will lose field association logic
chunkOverlap150–200 charactersEnsure that associated fields such as cross-chunk transaction remarks and amounts are not completely truncated
PARSE_OCR_ENABLEEnabledMost funding source documents are paper scans printed by banks, requiring OCR to extract text
UPLOAD_FILE_MAX_SIZE200 MBAnnual bank statement files are usually large, so large file upload support is needed
RECALL_TOP_KTop 8–12 entriesFunding source verification requires associating multiple transactions. Too many recalled entries increase computational load, while too few will miss key transactions
PARSE_TABLE_STRATEGYRetain complete table structureTransaction details for funding sources are mostly presented in tables. Splitting tables will destroy the correspondence of transaction fields

The parameter values provided on this page are conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require analysis based on individual cases, and it is recommended to test on your own samples before finalizing the settings.

Three common mistakes

  • When uploading large annual statement PDFs, an "offset out of range" prompt appears after the upload progress reaches 90%. The UPLOAD_FILE_MAX_SIZE parameter was not adjusted, and the default threshold is insufficient for a single annual statement file.
  • A large number of duplicate transaction record fragments appear after chunking, making it impossible to accurately match KYC verification requirements. No reasonable chunkOverlap interval was configured, or the chunk length was set incorrectly, causing the same text segment to be split and included multiple times.
  • After uploading a structured funding source document converted from Excel to PDF, the chunking result fails to extract complete transaction detail fields. The PARSE_TABLE_STRATEGY was not set to retain complete table structure, causing the table to be split line by line and losing field correspondence.

How to confirm the configuration is correct

  • Upload one typical funding source document, such as a single-month bank statement PDF, and check the parsed chunked text to confirm that core fields including transaction date, amount, and transaction counterparty are not split across chunks.
  • After adjusting configuration parameters, upload multiple funding source documents of different formats, such as scanned documents, native PDFs, and Excel-exported PDFs, to verify the completeness of OCR recognition and table parsing.
  • Initiate a KYC verification request, check whether the recalled chunked content covers the required transaction details, and confirm that the number of recalled entries meets the verification requirements.
  • Upload a single large annual statement file to verify that the upload progress completes normally without offset errors.

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.