What the data for this category looks like
Data sources include wealth management consultation script libraries from financial institutions, insurance product marketing materials, customer service standard copy, and event promotion assets. Update rhythm aligns with marketing campaign cycles. Batch updates occur when new campaigns launch, and only minor tweaks to existing content are made on a daily basis. Document structure includes product associated ID, marketing scenario tag, applicable customer group scope, copy body, delivery channel identifier, and effective time period. Most individual documents range from hundreds to thousands of characters, stored grouped by marketing batches.
What constraints these characteristics impose on vector models and indexing
The storage structure grouped by marketing batches requires indexes to support filtering and recall by batch ID metadata, to avoid retrieving irrelevant content across batches. High-frequency batch updates and daily minor tweaks tied to campaign cycles require indexes to support incremental synchronization, eliminating the need for full index reconstruction each time. The presence of multi-field metadata (scenario tags, customer group scope) requires indexes to support multi-dimensional combined filtering to narrow recall scope. The large variation in individual copy length requires vector models to adapt to variable-length inputs, while chunking strategies must match the information density of different marketing content.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Matches the maximum context window of most localized vector models, avoids overloading single segments with text information, and adapts to the typical length of comprehensive service marketing copy |
vector_db_retrieve_topk | Top 10–15 results | Comprehensive service marketing content has a large number of materials per batch, requiring coverage of enough scenario-related content while avoiding redundant recall that impacts response speed |
embedding_model | Prioritize Chinese open-source models such as bge-large-zh-v1.5 or m3e-base | Financial marketing content is primarily in Chinese, and local deployment avoids data compliance risks while adapting to the inference capabilities of local hardware |
index_partition_field | Marketing batch ID | Isolating indexes by marketing batch allows quick filtering of content from unrelated batches, improving retrieval efficiency and result accuracy |
embedding_batch_size | 32–64 entries | Adapts to hardware configurations with 32GB of memory, preventing out-of-memory errors during batch embedding |
similarity_threshold | 0.72–0.85 | Scenario similarity of comprehensive service marketing content is relatively high, requiring a reasonable threshold to filter low-relevance retrieval results |
The parameter values provided on this page are all conventional recommendations used as a starting point 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 configuration mistakes
- Phenomenon: Vector retrieval returns no results or far fewer results than the configured recall limit. Cause: Metadata filtering rules are not configured, causing the retrieval scope to cover unrelated marketing batch content.
- Phenomenon: Batch embedding tasks trigger out-of-memory errors, with
out of memoryappearing in logs. Cause: The batch embedding batch size exceeds the memory capacity of the hardware, and does not match the currently deployed hardware configuration. - Phenomenon: Knowledge base capacity statistics deviate significantly from the total size of original documents. Cause: Vector storage usage is estimated using the total character count of original documents directly, rather than the total character count of chunked text.
How to confirm correct configuration
- Upload a single typical marketing copy, check the chunked text blocks generated in the backend, and confirm that chunk lengths fall within the
chunk_sizeconfiguration range. - Select documents from a specified marketing batch to initiate retrieval, verify that only content associated with that batch is included in recall results.
- Run a batch embedding task, monitor hardware memory usage, and confirm no out-of-memory errors are triggered.
- Input a test query matching a specific marketing scenario, adjust the similarity threshold, and verify that the relevance of retrieval results meets business expectations.
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.