What the data for this category looks like
Data sources include promotional materials exported from the in-bank marketing platform, customer communication scripts approved by the compliance department, electronic materials for offline events, and quarterly updated marketing plans. Update rhythm varies by material type: event materials adjust with activity cycles, product materials update with business launch schedules, and compliance scripts iterate per regulatory requirements. Document structures include fixed fields such as marketing topic, target customer group, compliance ID, effective date, and expiration date. Main formats are PDF, PPT, and Word; some include structured rate and quota information, with units primarily in ten thousand yuan, natural days, and date formats.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Mixed-format materials require the parsing module to support PPT text block extraction and PDF table structured conversion, to avoid losing structured rate and quota information. Fields with effective and expiration dates require the retrieval link to support date range filtering, to recall content within the current validity period and avoid sending expired marketing information to customer groups. Metadata fields for compliance ID and target customer group require support for metadata retrieval, combined with full-text retrieval to narrow the recall scope and improve accuracy. Frequently updated materials require the knowledge base sync frequency to match business rhythms, to avoid content lag.
How to Set Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Marketing materials for joint-stock commercial banks are mostly multi-page PPTs and PDFs with tables, so parsing time is longer than that for general documents |
maxChunkSize | 1000–1200 characters | Marketing content includes structured information and coherent scripts, so the segment length needs to balance semantic completeness and retrieval accuracy |
RECALL_TOP_N | Top 8 entries | Needs to cover marketing needs of different customer groups while controlling context window usage |
SIMILARITY_THRESHOLD | 0.75–0.85 | Marketing content has relatively high semantic similarity, so low-relevance results need to be filtered out to retain accurately matched content |
SYNC_INCREMENTAL_INTERVAL | Every 4 hours | Matches the update frequency of marketing materials, which ranges from weekly to daily, to avoid resource waste and content lag |
UPLOAD_FILE_MAX_SIZE | 500 MB | Supports uploading large marketing event materials containing high-definition assets |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require individual analysis, and it is recommended to test on your own samples before finalizing settings.
Three Common Mistakes
- After configuring a Feishu knowledge base, PPT and PDF documents cannot be synced, and the sync task shows a failure. Cause: The Feishu application has not been granted read permissions for the corresponding document types, or the Feishu documents have not been set to be accessible by the specified application.
- After upgrading to version 4.9.0, an error is triggered when creating a knowledge base via URL. The backend log returns
cannot fetch internal url. Cause: The network policy of the local deployment environment restricts access to internal URLs, or the resource pointed to by the target URL has authentication requirements. - When clicking chunk preview for some PDF documents in the knowledge base, the interface displays a prompt that "the file content cannot be read". Cause: The PDF contains encrypted content, non-standard typesetting, or a damaged file structure, so the parsing module cannot complete text extraction and chunking.
How to Confirm the Configuration Is Complete
- Upload a typical marketing material PDF, check if the parsed text blocks contain complete structured information and script content, to confirm that the parsing parameters are effective.
- Initiate a retrieval test, enter a query term related to customer group scripts, check if the effective dates of the recalled results are within the current range, to confirm that the metadata filtering configuration is effective.
- View the knowledge base sync logs, confirm that the incremental sync tasks execute at the set interval, with no timeout or failure records.
- Trigger multiple retrieval tests for different types of marketing content, compare the relevance and number of results, to confirm that the parameter configuration meets business needs.
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.