What the Data for This Category Looks Like
Water utility ledger and report data originates from multiple sources: real-time pressure and flow data from pipeline monitoring nodes, daily operation reports from water plant central control systems, inspection logs from drainage pump stations, user repair work orders, and scanned paper ledger documents. Update frequencies vary significantly: real-time monitoring data generates new records every second, work orders and inspection reports are aggregated daily, and monthly ledgers are updated once per month. The data structure mixes structured fields and unstructured attachments. Structured fields include pipeline ID, monitoring point name, pressure value (unit MPa), flow rate (unit m³/h), repair time, processing result, and other related items. Unstructured attachments include inspection photos, scanned handwritten ledgers, and Excel summary sheets.
Constraints Imposed on Vector Models and Indexing
Mixed data types require indexes to support vector encoding for text, numerical values, and images. Without this support, monitoring data and attachment information cannot be fully correlated. Data with different update frequencies require adapted indexing strategies: real-time data uses incremental indexing to avoid delays from full updates, while batch reports can trigger full indexing on a daily basis. Adjust the context window via defaultToken to meet the encoding needs of mixed data. Unit information in structured fields affects vector semantics. Uncleaned units will cause matching deviations in similarity queries. Long-text inspection reports and short-text work order records need tailored chunking rules to preserve complete semantics.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
defaultToken | 4096 | Matches the default context window of most open-source embedding models, supports semantic encoding of single water utility data entries |
maxToken | 8192 | Accommodates multiple linked data segments during monthly ledger aggregation, meets the generation requirements for long documents |
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of inspection reports and the total number of fields in work order records |
| Overlap Length | 100–150 characters | Prevents logical breaks in monitoring data across chunks, preserves the relationship between adjacent records |
| Similarity Threshold | 0.72–0.78 | Semantic similarity range that distinguishes abnormal water pressure fluctuations from normal operational data |
| Single Search Limit | Top 10 entries | Covers the number of monitoring nodes linked to a single ledger, meets the information needs of summary reports |
Three Common Configuration Errors
- Issue: After deploying FastGPT v4.17.0 locally, no image indexing model option is available when creating a new knowledge base. Cause: The commercial image embedding plugin is not enabled, and only text embedding models load by default.
- Issue: Linked data matching deviation in ledger summary reports is excessive. Cause: No pre-cleaning is performed for unit fields in water utility data. Unit information is lost during vector encoding, leading to semantic confusion.
- Issue: Index update delay for real-time monitoring data is too high. Cause: No incremental index trigger rules are configured, and the full index update process is used instead.
How to Confirm Correct Configuration
- Upload a single ledger document that includes inspection photos. Review the index task log to confirm that an image embedding task record exists.
- Enter a search query containing "pressure 0.2MPa". Verify that the matching fields in the returned results include monitoring data with the corresponding unit.
- Test pushing a single pipeline monitoring data entry in real time. Confirm that the update progress of the index queue is completed within the preset time.
- Access the knowledge base configuration panel to confirm that the enabled embedding models include both text and image types.
All recommended configuration values are common starting points. Adjust them based on your own specific data samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.