What This Category of Data Looks Like
Marketing content data for large banks comes primarily from a centrally managed official marketing material library. The library includes product prospectuses, credit product introductions, offline event announcements, standard customer service response scripts, and more. Data updates follow two schedules. Standardized product information updates quarterly alongside product iterations. Temporary event-based content updates monthly or on an ad-hoc basis. Individual documents use structured formats. They contain fixed fields including product name, target customer group, fee/interest rate, application channel, and risk disclosures. Corresponding units for these fields include percentages, ten thousand yuan, months, years, and others.
Constraints Imposed by These Characteristics on Vector Models and Indexing
Content with multiple structured fields and fixed units requires vector models to distinguish between semantics and field attributes. This avoids interference from unit text on matching accuracy. Content with varied update schedules requires indexes to support two modes: incremental synchronization and full reconstruction. This meets the rapid update needs of temporary events. Individual documents have large length differences, ranging from hundreds of words of event copy to thousands of words of product prospectuses. Segment configuration must adapt to text splitting for different lengths, to ensure semantic integrity.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
EMBEDDING_MODEL_NAME | text-embedding-ada-002 or locally deployed m3e-base | The semantic complexity of large bank marketing content is moderate. These models cover mainstream text encoding needs |
CHUNK_SIZE | 800–1200 characters | When split from individual marketing materials, this length preserves complete product information units and avoids semantic breaks |
INDEX_INCREMENTAL_SYNC | Enabled | Meets the rapid update needs of temporary event-based content, and reduces resource consumption from full index reconstruction |
SIMILARITY_THRESHOLD | 0.75–0.85 | Filters low-relevance recall results and avoids redundant content interfering with marketing content matching |
RECALL_TOP_K | Top 10 results | In large bank marketing scenarios, around 10 recall results cover the main matching items for user queries |
PARSE_FIELD_MAPPING | Split fields by product name, target customer group, core benefits, and risk disclosures | Structured field splitting improves the accuracy of vector recall and avoids interference from unrelated unit text on semantic matching |
The parameter values provided on this page are common starting points for configuration setup. Actual values are affected by material format, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Newly created knowledge base data indexes stay in "indexing" status for an extended period. The cause is that incremental synchronization configuration is not enabled. Full index processing handles a large volume of historical marketing materials, exceeding the system's default timeout threshold.
- An "embedding model not configured" error returns when calling the vector model. The cause is that the target vector model is not added to the open interface channel. Relevant model access parameters are not correctly filled in the configuration file.
- A large number of unrelated marketing content items appear in recall results. The cause is that field mapping splitting is not configured. Unit text such as amount and interest rate mixes during vector encoding, leading to deviations in semantic matching.
How to Verify Correct Configuration
- Navigate to the vector model configuration page. Confirm the target model's access channel is added, and the model name exactly matches the
EMBEDDING_MODEL_NAMEconfiguration item. - Upload a single test marketing material. Check the parsed field splitting results, and confirm the
PARSE_FIELD_MAPPINGconfiguration has taken effect. - Initiate a test query. Check that the number of recall results matches the
RECALL_TOP_Kconfiguration, and similarity scores fall within the set range. - Submit a temporarily modified marketing content item. Check that the index updates automatically, and confirm the incremental synchronization configuration has taken effect.
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.