🤖 Agentic AI Integration¶

BfxPM features a built-in AI Assistant designed to help bioinformaticians manage their projects, answer technical questions, and automate repetitive tasks. This system is powered by the SmolAgents framework and is optimized for both privacy and performance.

🌟 Overview¶

The BfxPM AI suite is not just a chat interface; it is a Project-Aware Agent. It understands the context of your research, the structure of your files, and the metadata of your project.

Key Capabilities:¶

Project Context: The agent can read your project_tracker.yaml, directory structure, and basic file contents to provide tailored advice.
Tool-Assisted: It has access to specialized tools for listing files, checking system resources, and reading documentation.
Safety-First: A dedicated Safety Interceptor prevents accidental data loss from destructive suggestions.
Transparent Reasoning: Watch the agent's logic in real-time with the "Internal Thoughts" panel.

AI Setup Guard

In bfxpm ai setup, if you select any cloud provider (Gemini, OpenAI, Anthropic, Groq, or Mistral), the tool now displays a ⚠️ Privacy Alert panel.
It informs the user that prompts and metadata may be processed outside the EU and requires an explicit yes/no confirmation to proceed.
Selecting Ollama remains the recommended path for 100% local, EU-compliant execution.

⚙️ Setup¶

Before you can use the AI agent, you need to configure your environment.

bfxpm ai setup

This command runs an interactive assistant. Note: You can type q, exit, or abort at any prompt to cancel the setup process immediately.

The assistant will prompt you to:

Choose a Provider: Select between:
- Ollama (Local): Recommended for privacy-first research.
- Google Gemini: Top-tier cloud intelligence via the new GenAI SDK.
- OpenAI (ChatGPT): Industry-standard models like GPT-4o.
- Anthropic (Claude): World-class reasoning with Claude 3.5 Sonnet.
- Groq: Llama-3 models at lightning-fast speeds.
- Mistral AI: European-built leading open-weight models.
API Key: Provide the key for your selected provider. BfxPM stores them securely in an internal dictionary so you can switch providers without re-entering keys.
Model Selection: Choose your preferred model (e.g., gpt-4o, claude-3-5-sonnet, llama-3.1-70b).

💬 Interaction Modes¶

There are two primary ways to interact with the BioAssistant.

1. The Interactive Chat (`chat`)¶

For complex, multi-step tasks or general research support.

bfxpm ai chat

Inside the chat, you can ask things like:

"What is the best way to organize my differential expression results?"
"Can you show me the current project structure and suggest improvements?"
"Explain the difference between BAM and CRAM formats."

2. Single Question Mode (`ask`)¶

For quick answers or single actions without entering a full chat session.

bfxpm ai ask "How many FASTQ files do I have in my raw data folder?"

🛡️ Safety & Privacy¶

The Safety Interceptor¶

One of the most powerful features of the BfxPM agent is the Safety Interceptor. If the agent suggests a command that could be destructive (such as rm, mv, or rmdir):

Automatic Backup: BfxPM creates a timestamped backup of the target directory in .bfxpm/backups/.
Double Confirmation: The CLI will pause and require a manual y/n confirmation from you before executing the command.

Privacy-First Execution¶

We understand the sensitivity of genomic research data.

Ollama Integration: By choosing the Ollama provider, you can run the entire AI stack locally on your own hardware. Your metadata, prompts, and project details never leave your machine.
Encrypted Keys: Your API keys are stored securely in your local user configuration and are never shared or uploaded.

📂 Developer's Note: SmolAgents¶

Under the hood, BfxPM uses a custom ReAct (Reasoning and Acting) loop implementation via smolagents. This allows the tool to maintain a minimal memory footprint while providing high-level reasoning capabilities. The BioAssistant is specifically tuned for the Python 3.13 environment and utilizes the latest google-genai SDK for lightning-fast responses.