Natural Language Command Tool/Agent

A command-line tool which integrates with terminal/shell and interprets natural language commands and executes them as system operations using a Local Language Model (LLM). This tool can also search the internet, fetch information from webpages, and analyze web content.

Features

• Natural language processing of system commands
• Internet search capabilities
• Web content analysis and extraction
• Browser-based rendering for JavaScript-heavy websites
• Interactive shell mode and command-line mode
• Context-aware command history
• Secure operation handling with user confirmations
• Local LLM integration
• Persistent context management
• Notification system
• RAG (Retrieval Augmented Generation) support for enhanced contextual responses
  • Create and manage document collections
  • Semantic search across indexed documents
  • Context-aware responses using local knowledge base

Requirements

• Python 3.7+
• Local LLM server running on http://localhost:1234/v1 (tested with LM Studio)
• An LLM model which supports tool usage (Recommended Model: Qwen2.5-14B-Instruct)
• An embedding model (Recommended: text-embedding-nomic-embed-text-v1.5-embedding)
• Required Python packages:
  • docker
  • setuptools
  • twine
  • pytest
  • aiohttp
  • psutil
  • beautifulsoup4
  • duckduckgo-search
  • nltk
  • playwright (for browser-based web page rendering)
  • chromadb
  • langchain
  • langchain_core
  • langchain_community
  • sentence_transformers
  • pypdf
  • jq
  • unstructured
  • pdfminer.six
  • pi_heif
  • unstructured_pytesseract
  • pdf2image
  • pytesseract
  • opencv-python
  • tesserocr
  • unstructured-inference (Has some issues on Ubuntu ARM 64)
• Includes ai_code_sandbox as a submodule to execute python code in a sandbox.

Installation

1. Clone the repository
2. Create a soft link for ai_code_sandbox:

cd bashai
git submodule update --init
ln -s ai-code-sandbox/ai_code_sandbox

3. Install dependencies:

pip install -r requirements.txt

Note: If you have an "externally-managed-environment" you might need to create a virtual environment for python, for your pip to work.

python3 -m venv .
source bin/activate

4. Install playwright browser depending on your platform and preference.

python3 -m playwright install chromium
python3 -m playwright install firefox
python3 -m playwright install webkit

5. Install Tesseract.

Linux:

sudo apt install tesseract

MacOS:

brew install tesseract

Details

1. Contexts and security configuration are maintained in ~/.bashai folder
2. Configs can be overridden by creating a config.json file under the ~/.bashai directory. Sample configs:

{
  "llm": {
    "api_url": "http://localhost:1234/v1",
    "model": "local-model",
    "temperature": 0.7,
    "max_tokens": 4096
  },
  "context": {
    "max_length": 4000,
    "max_age_hours": 24
  },
  "logger": {
    "level": 10
  },
  "browser": "webkit",
  "rag": {
    "provider": "huggingface",
    "model": "all-MiniLM-L6-v2",
    "data_directory": "~/.bashai/embeddings",
    "chunk_size": 500,
    "chunk_overlap": 50,
    "distance_threshold": 0.7
  }
}

3. Logs are generated in /tmp
4. Contexts are maintained per shell and are auto cleaned up if the shell is unused for a prolonged period of time. Context can also be manually cleared. See the "Clean Context" section below.

Usage

• Make sure LM Studio is running and the LLM and embedding model is loaded.

Interactive Shell Mode

python agent

Command Line Mode

python agent <your prompt>

OR

agent <your prompt>

if you add agent to the PATH and chmod +x it (recommended).

RAG Operations

Create a new collection from documents:

agent "create a RAG collection called 'docs' from the ./documentation folder"

Query using RAG context:

agent "using the 'docs' collection, explain how to configure the logging system"

List available collections:

agent "list all RAG collections"

Delete a collection:

agent "delete the RAG collection named 'docs'"

Clean Context

Clears the context associated with the current shell.

agent -- --clean

Examples

# As a shell
$ python agent
agent> create a new file called test.txt
agent> show me the last 5 lines of /var/log/syslog
agent> search the web for latest Linux kernel features
agent> analyze the content from https://example.com/article
agent> fetch content from spa-webapp.com using browser rendering
agent> create a RAG collection called 'python-docs' from ./python/docs
agent> using python-docs collection, explain the asyncio module

# As a command
$ agent "create a backup of my-file.txt"
$ agent "what is the current weather in San Francisco"
$ agent "summarize the main points from https://example.com/blog-post"
$ tail -100 /var/log/syslog | agent analyze the logs
$ agent "index my-project-docs/ as a RAG collection called 'project'"

Configuration

The tool can be configured via ~/.bashai/config.json. Here are the available options:

{
  "llm": {
    "api_url": "http://localhost:1234/v1",
    "model": "local-model",
    "temperature": 0.7,
    "max_tokens": 4096
  },
  "context": {
    "max_length": 4000,
    "max_age_hours": 24
  },
  "logger": {
    "level": 10
  },
  "browser": "webkit",
  "rag": {
    "provider": "huggingface",
    "model": "all-MiniLM-L6-v2",
    "data_directory": "~/.bashai/embeddings",
    "chunk_size": 500,
    "chunk_overlap": 50
  }
}

RAG Configuration Options

• provider: Embedding model provider (currently supports 'huggingface')
• model: The embedding model to use for document indexing
• data_directory: Where to store the embeddings and collections
• chunk_size: Size of text chunks for document splitting
• chunk_overlap: Overlap between consecutive chunks
• distance_threshold: Similarity threshold for retrieval (0-1)

Security

The tool includes security measures for potentially dangerous operations:

• User confirmation for destructive operations
• Process isolation
• Context cleanup
• Error handling

License

This project is licensed under the MIT License - see the LICENSE file for details.

Author

[Aayush Chawla] mail@aay.sh