AI TaskVector

---

Beta version! Use it at your own risk.

---

Core concept: AI agents creating and training AI agents for users' needs.

A low-code programming framework and GUI visual editor for AI assistants to help users solve business and engineering tasks by composing specific workflows from units/pipilines, and configure/perform training via GUI and chat conversation—minimizing hand-written code.

• Language agnostic graph: the Graph is capable of carring units written in any language.
• Native runtime: Python-based graph execution.
• External runtimes (workflow conversion compatibility): Node-RED, Pyflow, ComFy, n8n, etc. You can drop in an external workflow as is, modify and export back. Use the external runtime "roundtrip" feature for RL training.
• Offline local models (no external API is required)
• Sustainable memory and RAG knowledge base

Assistants (co-pilots)

• Workflow Designer to create/modify workflows, generate custom units (if allowed), make integrations.
• RL Coach to train/fine-tune models.

---

Quick start

1. Install (from repo root)

python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Open the Constructor GUI (Flet)

Desktop app: workflow graph (canvas), training config, run/test, and AI chat (Workflow Designer / RL Coach).

pip install -r gui/requirements.txt
python -m gui.main

• Workflow: Load or paste a process graph (Node-RED/PyFlow/n8n/YAML); edit on canvas; run workflow, report, grep, GitHub from chat.
• Training: Load/edit training config (goal, rewards, callbacks); run training or test a saved model.
• Chat: Talk to Workflow Designer (graph edits) or RL Coach (training config); edits are applied to graph or config.

3. Train from the command line (optional)

python runtime/train.py --config config/examples/training_config.yaml

Use --process-config for a custom process graph; use --checkpoint to resume. All behavior is driven by the config files the assistants (or you) produce.

4. Test a trained model

python scripts/test_model.py ./models/temperature-control-agent/best/best_model

For a visual tank demo and manual sliders (thermodynamic example):

python -m environments.custom.thermodynamics.water_tank_simulator --config config/examples/training_config.yaml --model ./models/temperature-control-agent/best/best_model

---

Docker

You can run the app (and optionally the Ollama LLM server) in Docker. The image includes the full stack: main app, RAG, Flet GUI, and units (e.g. web_search). Works with classic Docker (e.g. 2022) and newer BuildKit. If you hit No space left on device during build, free disk space or set TMPDIR or PIP_CACHE_DIR to a directory on a larger drive before running docker build.

Build and run with Docker Compose (app + Ollama)

From the repo root:

docker compose build
docker compose up

Then open the Flet GUI in your browser at http://localhost:8550. The app is configured to use the Ollama service automatically via OLLAMA_HOST.

Pull a model in Ollama (one-time):

docker compose exec ollama ollama pull llama3.2

Models are stored in a persistent volume (ollama_data).

Build and run the app image only

docker build -t ai-taskvector .
docker run --rm -p 8550:8550 -e FLET_WEB=1 -e FLET_SERVER_PORT=8550 ai-taskvector

Open http://localhost:8550. If Ollama runs on your host, point the app at it with:

docker run --rm -p 8550:8550 -e OLLAMA_HOST=http://host.docker.internal:11434 ai-taskvector flet run gui/main.py --web -p 8550

Environment variables

Variable	Description
OLLAMA_HOST	Ollama server URL (default: http://127.0.0.1:11434). In Compose, set to http://ollama:11434.
OLLAMA_MODEL	Default model name (e.g. llama3.2) when not set in GUI settings.
OLLAMA_API_KEY	Optional; for Ollama Cloud.

Files

• Dockerfile — Full install (main + RAG + Flet GUI + units); default command runs the Flet GUI.
• docker-compose.yml — App + Ollama service; Flet runs in web mode on port 8550.

Apply assistant edits (workflows, same units as in-app chat):

• Graph: gui.components.workflow_tab.workflows.core_workflows.run_apply_edits then run_normalize_graph on the result (workflow gui/components/workflow_tab/workflows/core_workflows/apply_edits_single.json), or run that workflow via runtime.run.run_workflow with initial_inputs for inject_graph, inject_edits, inject_origin. See scripts/test_assistants.py.
• Training config: gui.components.workflow_tab.workflows.core_workflows.run_apply_training_config_edits (workflow gui/components/workflow_tab/workflows/core_workflows/apply_training_config_edits_single.json). Generic runner: python -m runtime <workflow.json> --initial-inputs @inputs.json — see runtime/README.md.

Stack

Dependencies are split across the root requirements.txt (full stack), gui/requirements.txt (Flet UI on top of the base install), and optional extras. Below maps areas of the product to the notable libraries declared there (and in rag/requirements.txt for RAG).

Area	Notable libraries	Declared in
Core (schemas, configs, graphs)	Pydantic, PyYAML, NumPy, Pandas; scikit-learn, Matplotlib where analytics/plotting are used	requirements.txt
Runtime (workflows, units, servers)	FastAPI, Uvicorn (LLM inference / ASGI); Requests, websocket-client (HTTP/WS and external adapters)	requirements.txt
Training (RL)	PyTorch, Gymnasium, Stable-Baselines3 (with extras), TensorBoard; tqdm, rich (CLI progress/logging); asteval, rule-engine (reward formula DSL and rule evaluation)	requirements.txt
GUI	Flet, flet-code-editor (workflow/code views)	gui/requirements.txt (install after requirements.txt)
RAG	LlamaIndex (llama-index, Hugging Face embeddings, Chroma vector store), ChromaDB, sentence-transformers, Docling (PDF/DOC/XLS ingestion)	rag/requirements.txt (optional; pip install -r rag/requirements.txt)
Assistants / chat	ollama (Python client to a local Ollama server; install models with the Ollama app / CLI separately)	requirements.txt

Optional unit extras (not in the root file): units/web/requirements.txt — DuckDuckGo search, BeautifulSoup4, html2text, minify-html. units/semantics/requirements.txt overlaps with root (lingua-language-detector, markdown-it-py, Pygments) for offline language detection and markdown rendering in units.

Install order (typical): pip install -r requirements.txt, then pip install -r gui/requirements.txt, then optionally pip install -r rag/requirements.txt and the unit extras above if you need those features.

License

MIT — use and modify for your projects.