Releases: botAGI/AGmind
Releases · botAGI/AGmind
v2.4.0 — Open WebUI Optional, Dify as Default Frontend
What's New
Features
- Open WebUI → Optional: moved to compose profile
openwebui. Disabled by default, saving ~300 MB RAM. Enable withENABLE_OPENWEBUI=truein wizard. - Dify App as default frontend:
http://<IP>/now serves Dify directly (was Open WebUI). Full API routes (/api,/v1,/console/api,/files,/e/) on port 80/443.
Fixes (11 total)
- SSE streaming:
proxy_buffering offon all API locations — fixes chat streaming on port 80/443 - Security:
cap_dropon api/worker containers (were missingsecurity-defaults) - Nginx: Authelia markers cleaned when disabled; atomic
.envupdate - Dify images: fallback pinned to
1.13.3(was:main) - Wizard: removed duplicate Portainer/Grafana LAN access question
- Wizard: VRAM summary now shows actual vLLM allocation (~90 GB, was showing 53 GB)
- ETL:
ETL_TYPE=Unstructuredcase-sensitive fix for Dify 1.13+ - Installer:
models.shprocess group kill — fixes hang at Phase 8 after model load - Installer:
gpu-metrics.shadded to runtime scripts copy — fixes empty GPU Grafana dashboard - Uninstall: handles all compose profiles, Docker volumes, systemd services, CLI symlink
Docs
- README updated (RU + EN): architecture diagram, endpoints, benchmarks, ARM64/DGX Spark
Benchmarks (NVIDIA DGX Spark, GB10)
| Metric | Result |
|---|---|
| TTFT (streaming) | 183 ms |
| TPS (single) | 23-24 tokens/sec |
| TPS (3 concurrent) | 50 tokens/sec aggregate |
| Context | 65K tokens (fp8 KV cache) |
| Max concurrency | 45 parallel @ 65K |
| Memory footprint | ~95 GiB (48.5 weights + 41.7 KV + overhead) |
Tested On
- NVIDIA DGX Spark (GB10, 128 GB unified memory, arm64)
- 25 containers, all services healthy
- Full connectivity matrix: nginx → api → postgres/redis/weaviate/vllm — all verified
- Network isolation: sandbox properly isolated from backend services
Full Changelog: v2.3.0...v2.4.0
v2.3.0 — Plugin Stability & Security Rework
Highlights
Полная переработка Plugin Daemon, выравнивание с каноном Dify 1.13.3, новая модель безопасности.
Plugin Daemon — полный рефакторинг
- Конфигурация приведена к официальному docker-compose Dify
- Plugin Daemon наследует shared env через
<<: *shared-env - Таймауты:
PYTHON_ENV_INIT_TIMEOUT=900,PLUGIN_MAX_EXECUTION_TIMEOUT=900,PLUGIN_DAEMON_TIMEOUT=600 - Nginx
/e/proxy timeout: 60s → 600s (закрывает "Plugin 300s failed") - Volume path выровнен с каноном
Security — tiered model
- Dify app-сервисы (API, Worker, Plugin Daemon) →
logging-defaults(без cap_drop) - Инфраструктура (DB, Redis, nginx, мониторинг) → полный
security-defaults - SSRF proxy восстановлен для API/Worker (как в каноне)
Dify 1.13.3 alignment
versions.envвыровнен с Dify 1.13.3- Фикс admin init:
admin@agmind.local→admin@agmind.ai - TEI model hang fix
- Redis hardening:
REDIS_MAX_CONNECTIONS=20, lock cleaner
Other
- GPU health probe fallback
- Docker pull: native TTY progress
- README: архитектурная диаграмма + секция безопасности обновлены
- Lint CI удалён
Updated Components
| Component | Was | Now |
|---|---|---|
| Dify | 1.13.2 | 1.13.3 |
| Plugin Daemon | main | 0.5.3-local |
| Sandbox | 0.2.12 | 0.2.14 |
Full Changelog: v2.2.0...v2.3.0
v2.2.0 — Upstream Updates (2026-03-22)
What's Changed
6 компонентов обновлены до последних версий. Все протестированы на тестовом сервере, 0 rollback'ов.
Updated Components
| Component | Was | Now |
|---|---|---|
| Dify | 1.13.0 | 1.13.2 |
| Open WebUI | v0.5.20 | v0.8.10 |
| Weaviate | 1.27.6 | v1.36.6 |
| vLLM | v0.17.1 | v0.18.0 |
| Xinference | v0.16.3 | v2.3.0 |
| Grafana | 11.4.0 | 12.4.1 |
Full Changelog: v2.1.0...v2.2.0
v2.1.0 — Initial Stable Release
AGMind v2.1.0 — Initial Stable Release
Tested on Ubuntu 24.04 with NVIDIA RTX 5070 Ti. 24/24 containers healthy.
Included
- Dify 1.13.0 (API + Web + Worker + Plugin Daemon + Sandbox)
- Open WebUI v0.5.20
- vLLM v0.17.1 (CUDA 13.0)
- TEI cuda-1.9.2
- Ollama 0.6.2
- Weaviate 1.27.6 / Qdrant v1.12.1
- Grafana 11.4.0 + Prometheus + Loki monitoring stack
- Authelia authentication
- Backup/restore with retention
- Offline bundle support
Features
- One-command install with interactive wizard
- GPU auto-detection (NVIDIA)
- Component update with rollback
- DR drill testing
- SSH hardening
- Secret rotation