What the data for this use case looks like
Energy storage yield and market trend data primarily comes from grid dispatching systems, power trading platforms, and site-side SCADA collection systems. Update cadence follows two schedules: Daily yield reports are generated at a fixed time each day. Real-time charge-discharge market data updates every 15 minutes. Each daily report includes structured fields such as unique site identifier, grid connection voltage level, total daily charge-discharge volume, unit charge-discharge cost, daily revenue, and cumulative revenue. Total charge-discharge volume is measured in kilowatt-hours. Revenue is measured in Chinese yuan. Unit cost is measured in yuan per kilowatt-hour.
What constraints do these characteristics impose on deployment and upgrade?
Requirements for multi-data source access mean multi-channel data synchronization adaptation rules must be configured. Data sources for different sites are scattered, so compatibility with different interface formats and permission rules is required. Differences in update frequencies demand precise configuration of scheduled tasks and streaming update logic. Daily reports must be synchronized at a fixed daily time. Real-time market data must match the 15-minute update interval. Complex field structures require precise mapping of field indexes for knowledge bases and vector databases, to avoid field misalignment or loss after parsing. Pressure from growing data volumes requires advance planning of vector database sharding and expansion strategies, to avoid performance bottlenecks during future upgrades.
How to set the configurations
| Configuration Item | Recommended Value | Rationale for this value |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Energy storage yield daily reports typically include multiple sets of data from multiple sites, so parsing time is longer than that of general documents |
maxContext | 8000–12000 characters | Historical yield comparison data across multiple sites must be retained. A sufficiently large context window improves query accuracy |
RECALL_TOP_K | Top 10–15 entries | A single daily report covers multiple sites, so enough relevant documents must be recalled to cover the query scope |
RERANKER_TIMEOUT | 300 seconds | Re-ranking calculations for multiple recalled results involve comparison of multiple fields, so processing time is longer than that of general text re-ranking |
UPLOAD_FILE_MAX_SIZE | 2000 MB | Monthly/quarterly aggregated energy storage yield reports typically have larger file sizes than general office documents |
SCHEDULE_CRON | 0 1 8 * * * | Energy storage yield daily reports are typically generated at 7 AM local time each day. Knowledge base updates must be completed within one hour after generation |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material format, 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 mistakes
- Symptom: A
400 Bad Requesterror occurs when configuring a model channel. Cause: The exclusive field format for energy storage data is not adapted. Required fields such as site ID and charge-discharge volume are not included in model request parameters. - Symptom: Knowledge base query responses time out when deployed. Cause: The vector database sharding strategy is not adjusted based on the daily incremental energy storage data, leading to single shard data volume exceeding the carrying threshold.
- Symptom: The platform fails to start after upgrading
fastgpt-mcp-server. Cause: Version matching identifiers for the main program and plugin are not updated synchronously, leading to incompatible inter-service protocols.
How to confirm the configuration is correct
- Manually upload a real energy storage yield daily report document. Check if the parsed fields in the knowledge base fully match the source document, with no missing or misaligned fields.
- Initiate a query for the yield of a specified energy storage site. Verify that the number of returned results matches the configured value of
RECALL_TOP_K. - View the scheduled task log. Confirm that the knowledge base update task triggered after 8 AM each day executes successfully with no abnormal errors.
- Call the test interface of the custom reranker model. Confirm that FastGPT can normally send requests and obtain return results through the configured custom request address.
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.