Jido AI

Jido AI Framework

Overview

Comprehensive Elixir framework for building sophisticated AI agents and workflows. Extends Jido framework with LLM capabilities, advanced reasoning techniques, and stateful conversation management.

Source: https://github.com/agentjido/jido_ai (upstream)
Fork: https://github.com/layeddie/jido_ai (your fork)

When to Use

• Multi-agent orchestration (coordinate multiple specialized agents)
• Advanced reasoning workflows (Chain-of-Thought, ReAct, Tree-of-Thoughts, GEPA)
• Multi-provider LLM access (57+ providers)
• Complex problem-solving with stateful conversations
• Local LLM support (Ollama, LM Studio)

Key Capabilities

• ✅ 57+ LLM providers (OpenAI, Anthropic, Google, Mistral, local)
• ✅ Advanced reasoning strategies:
  • Chain of Thought (+8-15% accuracy)
  • ReAct (+27% accuracy)
  • Self-Consistency (+17.9% accuracy)
  • Tree of Thoughts (+74% accuracy)
  • GEPA (+10.19% accuracy)
• ✅ Structured prompts (EEx + Liquid templates)
• ✅ Tool integration (function calling with schema conversion)
• ✅ Conversation management (stateful, ETS storage)
• ✅ Context window management (token counting, truncation)

Quick Start

# mix.exs
{:jido_ai, "~> 0.5.3"}

alias Jido.AI.{Model, Prompt}
alias Jido.AI.Actions.ReqLlm.ChatCompletion

# Create model
{:ok, model} = Model.from({:anthropic, [model: "claude-3-5-sonnet"]})

# Create prompt
prompt = Prompt.new(:user, "What is the capital of France?")

# Get response
{:ok, result} = ChatCompletion.run(%{model: model, prompt: prompt}, %{})
IO.puts(result.content)

Integration with ai-rules

• Plan Mode: Use jido_ai for multi-agent exploration of codebase architecture
• Build Mode: Use jido_ai reasoning strategies for complex implementation problems
• Review Mode: Validate reasoning quality and consistency across agent workflows

Agent Roles

• Architect: Design multi-agent architectures using jido_ai
• Orchestrator: Coordinate multiple specialized agents with jido_ai
• Reviewer: Validate reasoning quality and agent behavior patterns

Tooling Integration

• Use jido_ai with anubis_mcp for agent + MCP server workflows
• Use jido_ai with swarm_ex for lightweight agent coordination
• Use jido_ai with codicil for reasoning + code understanding

Reasoning Strategies

Chain of Thought

Purpose: Step-by-step reasoning

Accuracy Gain: +8-15%

Use Case: Complex multi-step problems requiring systematic breakdown

Pattern:

defmodule MyApp.ChainOfThought do
  alias Jido.AI.Runners.ChainOfThought

  def solve(problem, config \\ %{}) do
    ChainOfThought.run(%{
      problem: problem,
      llm: config.llm,
      max_iterations: 5
    })
  end
end

ReAct

Purpose: Reasoning with tool use

Accuracy Gain: +27%

Use Case: Problems requiring external tools or APIs

Pattern:

defmodule MyApp.ReAct do
  alias Jido.AI.Runners.ReAct

  def solve(problem, config \\ %{}) do
    ReAct.run(%{
      problem: problem,
      llm: config.llm,
      tools: config.tools
    })
  end
end

Self-Consistency

Purpose: Multiple reasoning paths with voting

Accuracy Gain: +17.9%

Use Case: Uncertain problems requiring diverse perspectives

Pattern:

defmodule MyApp.SelfConsistency do
  alias Jido.AI.Runners.SelfConsistency

  def solve(problem, config \\ %{}) do
    SelfConsistency.run(%{
      problem: problem,
      llm: config.llm,
      paths: 3
    })
  end
end

Tree of Thoughts

Purpose: Tree search exploration

Accuracy Gain: +74%

Use Case: Problems requiring exploration of solution space

Pattern:

defmodule MyApp.TreeOfThoughts do
  alias Jido.AI.Runners.TreeOfThoughts

  def solve(problem, config \\ %{}) do
    TreeOfThoughts.run(%{
      problem: problem,
      llm: config.llm,
      max_depth: 3
    })
  end
end

GEPA

Purpose: Evolutionary prompt optimization

Accuracy Gain: +10.19%

Use Case: Problems requiring iterative refinement

Resources

• HexDocs: https://hexdocs.pm/jido_ai
• Upstream: https://github.com/agentjido/jido_ai
• User Guides: https://github.com/layeddie/jido_ai/tree/main/guides/user
• Developer Guides: https://github.com/layeddie/jido_ai/tree/main/guides/developer

Provider Configuration

Supported Providers (57+)

# OpenAI
config :jido_ai, Jido.AI.Providers.Anthropic,
  api_key: System.get_env("OPENAI_API_KEY"),
  models: [
    gpt_4o: "gpt-4o",
    gpt_4_turbo: "gpt-4-turbo"
  ]

# Anthropic
config :jido_ai, Jido.AI.Providers.Anthropic,
  api_key: System.get_env("ANTHROPIC_API_KEY"),
  models: [
    claude_3_5_sonnet: "claude-3-5-sonnet-20241022",
    claude_3_opus: "claude-3-opus-20240219"
  ]

# Google
config :jido_ai, Jido.AI.Providers.Google,
  api_key: System.get_env("GOOGLE_API_KEY"),
  models: [
    gemini_1_5_pro: "gemini-1.5-pro",
    gemini_1_5_flash: "gemini-1.5-flash"
  ]

# Local (Ollama)
config :jido_ai, Jido.AI.Providers.Ollama,
  base_url: "http://localhost:11434",
  models: [
    llama3: "llama3"
  ]

Prompt Templates

EEx Templates

defmodule MyApp.Prompts do
  alias Jido.AI.Prompt

  def system_prompt do
    Prompt.new(:system, """
    You are an expert Elixir/BEAM developer.
    Follow OTP best practices and TDD workflow.
    Use Domain Resource Action pattern.
    Prioritize code quality and test coverage.
    """)
  end

  def user_prompt(template \\ "") do
    Prompt.new(:user, """
    Question: {{question}}
    Context: {{context}}
    Instructions: {{instructions}}
    """)
  end
end

Liquid Templates

defmodule MyApp.LiquidPrompts do
  use Jido.AI.Prompt.Templates.Liquid

  def system_prompt do
    """
    {% if is_authenticated %}
      You are authenticated as {{user_name}}.
    {% endif %}
    
    {{content}}
    """
  end
end

Conversation Management

Stateful Conversations

defmodule MyApp.Conversation do
  use Jido.AI.Conversation

  def start(user_id) do
    Conversation.start(user_id)
  end

  def add_message(conversation_id, role, content) do
    Conversation.add_message(conversation_id, role, content)
  end

  def get_history(conversation_id) do
    Conversation.get_history(conversation_id)
  end

  def get_context(conversation_id, max_tokens \\ 8000) do
    Conversation.get_context(conversation_id, max_tokens: max_tokens)
  end
end

Best Practices

• Start with Chain of Thought for simple problems
• Use ReAct when tools are available
• Try Tree of Thoughts for exploration tasks
• Use Self-Consistency for uncertain problems
• Use GEPA for iterative refinement
• Monitor token usage and context window
• Choose appropriate LLM provider for task complexity
• Leverage conversation management for multi-turn interactions

Troubleshooting

• Verify API keys are set correctly
• Check provider configuration for correct models
• Monitor reasoning performance (accuracy, token usage)
• Test reasoning strategies on sample problems
• Verify ETS storage is available for conversations
• Review prompt templates for effectiveness