Behaviour for LLM provider adapters.

Implement this behaviour to add support for additional LLM providers.

Creating a Custom Provider

Any module implementing the chat/2 callback can be used as a provider. Here's a minimal example:

defmodule MyApp.LLM.Groq do
  @behaviour LangEx.LLM

  @impl true
  def chat(messages, opts \\ []) do
    api_key = opts[:api_key] || System.get_env("GROQ_API_KEY")
    model = opts[:model] || "llama-3.3-70b"
    tools = opts[:tools] || []
    formatted_tools = Enum.map(tools, fn
      %LangEx.Tool{name: name, description: desc, parameters: params} ->
        %{type: "function", function: %{name: name, description: desc, parameters: params}}
      raw -> raw
    end)

    # Make API call, return {:ok, Message.ai(...)} or {:ok, Message.ai(nil, tool_calls: [...])}
    # ...
  end
end

Then register it so model-string resolution and ChatModel.node work:

# In your application startup or config:
LangEx.LLM.Registry.register_provider(:groq, MyApp.LLM.Groq)
LangEx.LLM.Registry.register_prefix("llama-", :groq)

Or use it directly without registration:

MyApp.LLM.Groq.chat(messages, model: "llama-3.3-70b", api_key: "...")

# In a graph node:
LangEx.LLM.ChatModel.node(provider: MyApp.LLM.Groq, model: "llama-3.3-70b")

Using LangEx.Config for Custom Providers

To use the shared config resolution (env vars, app config, explicit opts):

# In config/runtime.exs:
config :lang_ex, :providers,
  groq: %{env_key: "GROQ_API_KEY", default_model: "llama-3.3-70b"}

# Then in your adapter:
api_key = LangEx.Config.api_key!(:groq, opts)
model = LangEx.Config.model(:groq, opts)

Tool Calling

All built-in adapters support tool calling via the :tools option:

• :tools - list of %LangEx.Tool{} definitions. Each adapter translates to its native format by accessing the struct fields directly.

When the model requests tool calls the adapter returns {:ok, %Message.AI{tool_calls: [...]}}. Use LangEx.Tool.Node in your graph to execute tool calls and feed results back to the LLM.