What Data for This Category Looks Like
Funding source KYC data mainly comes from materials such as corporate current account statements, personal bank transfer receipts, tax filing forms, and income certificates. There are two update cycle types: enterprise customers sync updates quarterly or annually, while individual customers trigger updates when opening accounts or completing large transactions. Each single document includes structured fields such as counterparty account name, transaction amount, transaction time, fund purpose, and voucher number. It also includes attachments of original vouchers in PDF or scanned format. Structured fields uniformly use standard units: amounts are in Chinese yuan, and time follows the ISO 8601 format.
What Constraints Do These Characteristics Impose on Vector Models and Indexing
First, data includes structured metadata and unstructured original vouchers. This requires vector models to support encoding both text and image content, and indexes to support multimodal vector concatenation. Second, there are both batch and real-time update scenarios. Indexes must support incremental write and delete operations to avoid performance loss caused by full index reconstruction. Third, transaction amount is a standardized numeric field. Normalization must be completed during the preprocessing stage to ensure a reasonable vector space distribution. The presence of multiple fields requires prioritizing core fields such as transaction purpose and counterparty information for index construction, to avoid reduced retrieval efficiency caused by dimensional redundancy.
How to Configure
| Configuration Item | Suggested Value | Rationale for This Value |
|---|---|---|
enable_ocr | Enabled | Most original vouchers are PDFs or scanned documents, so OCR is required to extract text content for vector encoding |
chunk_size | 800–1200 characters | Structured descriptions and voucher text for funding source documents have moderate length. This segment range balances context completeness and vector dimensionality |
vector_store_batch_size | 50–100 items per batch | Batch ingestion reduces API call frequency, and adapts to batch update scenarios for funding source data |
recall_top_k | Top 10 items | KYC verification needs to cover multiple related transactions. A 10-item recall volume balances retrieval completeness and response speed |
similarity_threshold | 0.75–0.85 | Low-similarity irrelevant transaction records must be filtered. This range matches the business matching accuracy requirements for funding source verification |
metadata_embedding_enabled | Enabled | Metadata such as transaction amount and time can supplement vector semantic information to improve retrieval accuracy |
The parameter values provided on this page are general recommendations used to establish a starting point 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
- Calling the batch ingestion API returns a
400 Bad Requesterror with the promptinvalid batch size. The ingestion request is not split according to the batch size configured invector_store_batch_size, exceeding the single-submission limit allowed by the API. - The retrieval results return multiple duplicate transaction records, and the number of recalled items is far lower than the configured value. A separate index is created for each document instead of using a unified index to store the same type of funding source data, leading to cross-index retrieval failing to cover all data.
- The external management system cannot trigger add or delete operations on the vector database, and cannot create operation accounts with non-root permissions. The native vector database permission system is used directly, and permission encapsulation is not implemented through platform API interfaces, making it impossible to connect to the permission control requirements of the external management system.
How to Confirm the Configuration Is Correct
- Upload a single funding source voucher, check if the parsed text content and OCR result match the original voucher, confirming that the
enable_ocrconfiguration is active. - Submit a batch ingestion request, check the vector database write logs to confirm that the ingestion batch size matches the
vector_store_batch_sizeconfiguration. - Initiate a retrieval request, verify that the number of recalled results matches the
recall_top_kconfiguration value, and that the similarity scores fall within the preset range. - Initiate add or delete operations on the vector database through the external management system, verify that the index update process can be triggered normally, confirming that the permission encapsulation is active.
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.