This is the Jarvis Offline project - an AI agent system built with Python that uses Ollama for local LLM inference and various tools for enhanced functionality.
- Main agent class for interacting with Ollama API
- Supports conversation history, system instructions, and tool integration
- Features:
- Agent naming and role management
- Conversation context from other agents
- JSON schema support for structured responses
- Streaming and non-streaming responses
- Tool/function calling capabilities
- Contains utility functions and external API integrations
- Currently includes:
tavily_search()
: Web search using Tavily API with XML-formatted results- Environment variable loading with python-dotenv
- Use type hints for all function parameters and return values
- Follow PEP 8 naming conventions
- Include comprehensive docstrings with Args and Returns sections
- Use descriptive variable and function names
- Implement proper exception handling for external API calls
- Return meaningful error messages in XML format for tools
- Use timeouts for network requests (default: 30 seconds)
- Each agent should have a unique name and role
- Use conversation history for context continuity
- System instructions should remain separate from agent context
- Tool results should be captured and included in agent context
- Tools should return string responses
- Use XML tags for structured output (
<chunk 1>
, <error>
, etc.)
- Include metadata in tool responses
- Validate input parameters and provide clear error messages
- Use
.env
files for API keys and sensitive configuration
- Provide
.env.template
for easy setup
- Load environment variables at module level with python-dotenv
<agent_name>
Task: This is the role_description
Agent Response: response_text
Tool(s) Invoked and the tool outputs obtained: {tool_name: result}
</agent_name>
- System instructions set agent personality
- Conversation history maintains context
- Agent context tracks specific agent actions
- Tool results are integrated into responses
- ollama: Local LLM inference
- requests: HTTP client for API calls
- pydantic: Data validation and JSON schema
- python-dotenv: Environment variable management
- Include test functions in
if __name__ == "__main__":
blocks
- Test error conditions and edge cases
- Verify API integration without requiring actual API keys in tests
- Use descriptive output for debugging
def tool_name(param1: str, param2: int = 5) -> str:
"""
Brief description of what the tool does.
Args:
param1 (str): Description of parameter
param2 (int): Description with default value
Returns:
str: Description of return value with XML format
"""
try:
# Implementation
return "<chunk 1>Result</chunk 1>"
except Exception as e:
return f"<error>{str(e)}</error>"
agent = OllamaClient(
model_name="llama3.2:3b",
system_instructions="You are a helpful assistant...",
agent_name="AgentName"
)
agent.set_role("Specific role description")
jarvis-offline/
├── ollama_client.py # Main agent implementation
├── tools.py # Tool functions and utilities
├── .env.template # Environment variable template
├── .env # Actual environment variables (gitignored)
├── requirements.txt # Python dependencies
└── README.md # Project documentation
- Consider impact on existing conversation flows
- Maintain backward compatibility where possible
- Add appropriate type hints and documentation
- Include error handling and validation
- Update this instructions file if adding new patterns