What the data for this category looks like
Data sources for water utility marketing content include official public account posts from water service operators, offline marketing material copy, regional water supply service announcements, and water conservation science popularization content materials. Updates are triggered by marketing campaigns or service adjustments, with no fixed cycle. Single-update document lengths range from hundreds to thousands of characters. Document structures include fields such as title, release time, service coverage area, event details, and contact information. Some content includes unit annotations for service area in square kilometers and number of served households. Some materials also include alt text descriptions for real-world images.
What constraints do these characteristics impose on vector models and indexing
The non-fixed update rhythm and varying text lengths of water utility marketing content require vector models to support dynamic incremental indexing and adaptive chunking strategies. Content contains numerical fields with units, such as service area and number of served households. Vector modeling must retain the semantic association between numerical values and their units, to avoid losing meaning after splitting. Marketing content has strong timeliness: effective cycles for event announcements and price adjustment notices are limited. Indexes must support fast filtering and recall based on release time and event cycles, as well as batch cleanup of expired content. Some content includes image alt text, which must be associated with image feature vectors to ensure consistency in cross-modal text-image retrieval.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Balances chunking for varying-length water utility marketing content, preserves semantic integrity of key information such as service area and event details |
chunk_overlap | 100–150 characters | Prevents contextual breakage after chunking, retains association information between numerical values and units across chunks |
similarity_top_k | Top 8–12 results | Adapts to the regional service attribute of water utility marketing content, controls single recall scope to reduce irrelevant content interference |
similarity_threshold | 0.72–0.80 | Filters low-relevance generic copy, retains retrieval results strongly linked to regional services and event themes |
index_update_strategy | Incremental update mode | Adapts to the non-fixed update cycle of marketing content, reduces resource consumption from index rebuilding |
IMAGE_EMBEDDING_ENABLE | Enabled | Synchronously generates feature vectors for marketing material images, supports cross-modal text-image retrieval |
The parameter values provided on this page are common recommended 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 mistakes
- Phenomenon: When identical water utility content retrieval requests are sent repeatedly, the system continuously returns vector database query logs. Cause: No query cache-related parameters are configured, and duplicate request cache reuse logic is not enabled.
- Phenomenon: After importing marketing documents using "Add Data to Collection", some content containing service area and household count cannot be accurately recalled. Cause: No reasonable
chunk_overlapparameter is set, causing loss of association information between numerical values and units across chunks. - Phenomenon: After calling the "Create Training Order" interface, no usable knowledge base index is generated. Cause: Automatic training trigger configuration is not enabled, or the vector database instance is not correctly bound, causing training tasks to fail to execute.
How to confirm correct configuration
- Send a retrieval request containing regional service keywords, wait for the cache to expire, then send the same request again. Verify whether cached results are reused.
- Import a marketing document containing numerical units, then check during retrieval whether the returned chunks retain the associated semantics of numerical values and units.
- Submit a training order task, check whether the corresponding index file is generated in the vector database console, and confirm that the index update status is normal.
- Upload a matching marketing material image, send a retrieval request containing keywords from the image alt text, and confirm that the image can be recalled synchronously.
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.