Details of FastGPT v4.16.2 Technical Improvements
UI & Administrative Enhancements
This release includes targeted UI and routing updates for end users and administrators. The Account page UI and mobile responsiveness have been fully improved. Administrative routing for Admin pages and model providers has been optimized. Tool configuration, version selection, and parsed-document views now use responsive layouts, two-column grids, and ellipsis tooltips for long filenames. Payment modals update immediately when switching payment methods, and invoice submission refreshes bill and invoice records in-place without reloading the entire page.
Reliability & Error Standardization
Key reliability and error handling improvements are summarized below:
| Improvement Category | Specific Changes |
|---|---|
| Upload Reliability | Enhanced S3 chunked upload reliability on unstable networks; standardized all file upload error messages |
| Tool Call Anomaly Detection | Flags 5+ consecutive calls to the same tool with identical arguments as a model hallucination |
| Failed Tool Logging | Preserves failed tool responses from Agent calls, exposing failures via nodeResponse, toolResponse, and streaming events |
| Workflow Billing | Unifies sub-workflow output and billing across Workflow Tool, ToolCall, and Agent runs. Commercial Workflow Tool failures no longer incur call or token fees, while pluginOutput.error from personal Workflow Tools remains a dedicated business field |
Performance & Resource Optimization
Several backend and performance updates reduce operational overhead and risk: Skill saving has been refined by extracting name and description as structured fields, with description now optional. Administrative queries for teams, plans, payments, and user lists have been improved with username search support and faster pagination performance. PDF parsing charges are now attributed to the read_files tool node, with page counts preserved. The usageId field is propagated through external-file and API-file Dataset parsing to avoid orphaned usage charges. Milvus vectors and full-text data are stored in the same collection, with full-text granularity aligned to indexes[].text; image vectors do not index image URLs or object storage keys. File parsing memory and concurrency controls have been updated: queued tasks now use only a lightweight file source instead of full Buffer objects. Trusted S3 files reserve resources based on their known size, while external URLs update resource reservations as bytes stream in, with hard limits on file size and per-task memory to reduce out-of-memory (OOM) risk during concurrent parsing. Unified external-file download and parsing across Datasets, Chat, and Workflows ensures consistent handling, with uploaded files (including Dataset templates, backups, and App evaluation CSVs) sent to S3 before parsing to avoid retaining multipart temporary files in queues.