# GptAgent

GptAgent is an Elixir-based service that provides a conversational agent
interface using the OpenAI GPT models. It allows for the integration of
GPT-powered conversations within various platforms and services.

## Installation

If [available in Hex](, the package can be installed
by adding `gpt_agent` to your list of dependencies in `mix.exs`:

def deps do
    {:gpt_agent, "~> 2.0"}

Documentation can be generated with
[ExDoc]( and published on
[HexDocs]( Once published, the docs can be found at

## Configuration

To configure the GptAgent, you need to set the following environment variables
in your project's runtime config file (usually `config/runtime.exs`):

config :open_ai_client, :base_url, System.get_env("OPENAI_BASE_URL") || ""
config :open_ai_client, :openai_api_key, System.get_env("OPENAI_API_KEY") || raise("OPENAI_API_KEY is not set")
config :open_ai_client, :openai_organization_id, System.get_env("OPENAI_ORGANIZATION_ID")
config :gpt_agent, :heartbeat_interval_ms, if(config_env() == :test, do: 1, else: 1000)
config :gpt_agent, :timeout_ms, get_env("GPT_AGENT_RECEIVE_TIMEOUT_MS", 120_000, :int)

# if this is set to a number larger than the `:timeout_ms` above, the above value will be used instead
config :gpt_agent, :receive_timeout_ms, get_env("GPT_AGENT_RECEIVE_TIMEOUT_MS", 30_000, :int)

Make sure you have the `OPENAI_API_KEY` and (optionally) `OPENAI_ORGANIZATION_ID`
system environment variable set to the correct values for your API key and OpenAI

You can also configure the logger level in the config file:

config :gpt_agent, :log_level, :warning

This will set the logging level to `:warning`, which is the recommended level
for production. You can change it to `:debug` for more verbose logging during

## Usage

First, ensure the supervisor is running in your applications supervision tree.
For example:

children = [
  {GptAgent.Supervisor, []}

opts = [strategy: :one_for_one, name: MyApp.Supervisor]
Supervisor.start_link(children, opts)

Before you can start an agent, you need to have both a thread_id and an
assistant_id from the OpenAI assistants API.

To create an example assistant, you can run `"/v1/assistants",
json: GptAgent.Assistants.MemGpt.schema())` and note the id returned in the
request body.

```elixir"/v1/assistants", json: GptAgent.Assistants.MemGpt.schema())
#=> {:ok, %Req.Response{ body: %{ "id" => "asst_1Ut1Wxnw0MQAF5G3qWcoMRIQ", ...}, ...}

You can create a new thread with:
 {:ok, thread_id} = GptAgent.create_thread()

Then you can start the agent with:

{:ok, pid} = GptAgent.connect(thread_id: thread_id, assistant_id: assistant_id)`

This will start the agent process if one is not already running for the thread.
We use a process registry to ensure that there is only one agent process running
for a given thread_id. It will also use `Phoenix.PubSub` to subscribe the
current process to messages published by the thread's process. Expect to receive
the following messages:

  - `%GptAgent.Events.UserMessageAdded{}`: Triggered when a user message is
    added to the thread.
  - `%GptAgent.Events.RunStarted{}`: Indicates that a run has been started for
    the thread.
  - `%GptAgent.Events.ToolCallRequested{}`: Occurs when a tool call is
  - `%GptAgent.Events.ToolCallOutputRecorded{}`: Recorded when the output of a
    tool call is captured.
  - `%GptAgent.Events.RunCompleted{}`: Signifies that a run has been completed.
  - `%GptAgent.Events.AssistantMessageAdded{}`: Triggered when an assistant
    message is added to the thread.

To add a user message to the thread and run it with the default assistant: `:ok
= GptAgent.add_user_message(pid, "Hello, world!")`

You must at the very least monitor for the `ToolCallRequested` events (if your
assistant uses tool calls), so that you can submit the results back to the run.
To submit the results, use `GptAgent.submit_tool_output(pid, tool_call_id,

### Timeouts

The GptAgent processes will, by default, stay running for 2 minutes, after
which, if no additional activity has taken place, the process will shutdown
normally. If you would like to set a different timeout value, you can pass the
`timeout_ms` option to `GptAgent.connect/1`:

# Shut down the process if it has not received any activity within 200ms
{:ok, pid} = GptAgent.connect(thread_id: thread_id, assistant_id: assistant_id, timeout_ms: 200)

There is also a receive timeout value for the HTTP requests to OpenAI. This is
specified via the `:receive_timeout_ms` configuration value in your application
config, however if the agent timeout is set to a smaller value, the agent
timeout will also be used as the receive timeout. This is to prevent situations
where the Agent will hang due to OpenAI taking too long to respond.