What data for this category looks like
Water treatment research and investment data sources include environmental protection department water quality monitoring bulletins, industrial water treatment operation logs, third-party testing institution reports, water treatment equipment manufacturer parameter manuals, and municipal pipe network operation records. Four update rhythm types exist: real-time monitoring data (second to hour level), monthly water quality reports (updated monthly), static equipment parameters (updated with upgrades), and sudden pollution emergency data (temporarily added). Document structures include dozens of pages of annual analysis reports, structured monitoring tables, and single indicator records. Common fields and units include COD (mg/L), ammonia nitrogen (mg/L), pH (dimensionless), flow rate (m³/h), pressure (MPa), and others.
Constraints imposed by these characteristics on knowledge base retrieval and recall
Multi-source heterogeneous data structures require retrieval systems to support differentiated synchronization strategies. Distinct trigger rules apply for incremental and full synchronization. Mixed document types of long text and structured tables require dual adaptation. The system must support long text segment parsing and table field extraction. This avoids missing associated indicator information. Indicator fields with multiple units can interfere with similarity calculation logic. Additional unit verification is needed. This prevents misjudgments caused by matching only numerical values. Data sources with different update frequencies require corresponding synchronization cycles. This ensures retrieved data timeliness meets research and investment needs.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Water treatment documents include long-text analysis reports and structured tables. Excessively long segments lose indicator association context. Excessively short segments damage the integrity of single sets of monitoring data. 800–1200 characters cover the complete description of a single set of core indicators |
parse_table_enable | Enabled | The water treatment knowledge base contains many structured tables for water quality monitoring and equipment operation. Enabling this setting automatically extracts table fields and corresponding values, avoiding missing structured information during retrieval |
similarity_threshold | 0.65–0.85 | Water treatment indicators require matching both numerical values and units. A threshold that is too low introduces irrelevant monitoring data. A threshold that is too high misses similar indicators. Combining unit verification narrows the reasonable value range |
recall_top_k | Top 10–15 results | Water treatment research and investment require association of multiple linked indicators, such as pollution trend analysis of COD and ammonia nitrogen. 10–15 recall results cover multi-dimensional associated data |
rerank_top_n | Top 5 results | Research and investment reports require precise matching of core conclusions. Retaining the top 5 results after reranking filters low-relevance marginal monitoring data and focuses on core analysis content |
sync_interval | Configured by data source type: real-time synchronization / daily synchronization / monthly synchronization | Use real-time synchronization for real-time monitoring data, daily synchronization for monthly water quality reports, and monthly synchronization for static equipment parameters. This adapts to the update rhythms of different data sources |
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 specific analysis. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Phenomenon: Knowledge base search tests return scores exceeding 1000. Cause: Unit matching verification is not enabled, and only numerical values are matched, causing abnormal similarity calculation.
- Phenomenon: The interface becomes unresponsive after entering the agent editing page. Cause: The knowledge base contains ultra-large volume real-time monitoring log files. The
UPLOAD_FILE_MAX_SIZElimit is not set, causing loading timeout. - Phenomenon: Uploaded complex water treatment process tables cannot be parsed correctly. Cause: The
parse_table_enableconfiguration is not enabled, and tables are not split into single indicator entries.
How to confirm the configuration is complete
- Upload a structured table containing COD and ammonia nitrogen monitoring data, and check whether the analysis result correctly extracts fields and corresponding values.
- Enter preset research and investment query terms, and check whether the recall results include associated multiple sets of indicator data, and the scores are within a reasonable range.
- Configure synchronization intervals for different data sources, and check whether the knowledge base update log completes data synchronization according to the set cycle.
- Add a knowledge base search node to the workflow, and test whether the retrieval interface can be called normally after passing the preset knowledge base ID.
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.