What This Category’s Data Looks Like
Financial report data for the water utility industry comes primarily from public periodic reports, monthly operation reports, and industry disclosure materials published by listed water utility enterprises. The core update cycle is quarterly. Quarterly operation data is disclosed 1 to 2 months after the end of each quarter. Annual reports are published within the first quarter of the following year. Document structures usually include split business segment data, covering modules such as water supply, sewage treatment, and recycled water utilization. Fields include treatment scale, average daily water supply, unit operation cost, revenue breakdown and more. Common units are cubic meters, yuan per cubic meter, ten thousand yuan, and others.
Constraints Imposed on Vector Models and Indexing
The high-frequency updates and multi-module structured features of water utility financial reports impose multiple constraints on the vector models and indexing link. First, the high-frequency quarterly and monthly update requirements require indexing to support incremental update logic, avoiding computational resource waste and retrieval delays caused by full reconstruction. Second, a single document contains heterogeneous data from multiple business segments, and field units vary. Vector models need to support mixed encoding of multiple fields to eliminate semantic drift caused by unit differences. In addition, structured data with many detailed dimensions requires indexing to support precise filtering by report period and business segment, improving retrieval targeting.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
embedding_model | Use Doubao-embedding-large first | Adapts to the semantic encoding needs of structured text and field units in water utility financial reports |
index_chunk_size | 800–1200 characters | Matches the average length of single-segment structured data in water utility financial reports, avoiding semantic fragmentation caused by overly short chunks |
vector_recall_topk | Top 6–10 results | Covers retrieval needs for multiple business segments in water utility financial reports, preventing missed recall of data from specific segments |
similarity_threshold | 0.72–0.80 | Distinguishes semantic differences between similar business segments in water utility financial reports, reducing false recall rates |
enable_incremental_index | Enabled | Adapts to high-frequency quarterly and monthly update scenarios, reducing resource consumption from full index reconstruction |
enable_table_multi_vector | Enabled | Supports independent vector encoding of multi-module table data in financial reports, improving structured retrieval accuracy |
The parameter values provided on this page are general recommendations used as a starting point for configuration. Actual values are affected by material form, 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 Configuration Mistakes
- Issue: When enabling the
Doubao-embedding-largeindexing model in the knowledge base settings, filling in a custom request address and APIKey, then clicking test triggers an error. Cause: The format of the custom request address was not verified, or the APIKey does not have permissions to call the corresponding vector model. - Issue: Setting an overly small value for
index_chunk_size, causing structured financial report data to be split excessively, resulting in semantically incomplete fragments during retrieval. Cause: The actual length of single-segment data in water utility financial reports was not matched, and associated semantics between modules were lost after splitting. - Issue: Not enabling
enable_table_multi_vector, resulting in only a single vector being generated for table data in financial reports, making it impossible to accurately recall table content for specific business segments. Cause: Multi-vector support was not enabled for the multi-module table characteristics of water utility financial reports, leading to insufficient encoding accuracy for structured data.
How to Verify Proper Configuration
- Navigate to the vector model configuration page, verify the selection of
embedding_modeland the filled content of the custom request address and APIKey, confirming they match the configuration requirements of the target model. - Upload a test segment of a water utility financial report, check the parsed segment length, confirming that the value of
index_chunk_sizematches the actual data length. - Initiate a retrieval test, enter a query term related to a business segment, check the module coverage of the recall results, confirming that the values of
vector_recall_topkandsimilarity_thresholdmeet the scenario requirements. - View the incremental update logs of the knowledge base, confirm that newly added financial report data can automatically trigger index updates, verifying that the configuration of
enable_incremental_indextakes 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.