DORA is an open-source project implementing a dataflow-oriented middleware, designed to simplify and streamline the creation of AI-based robotic applications. It provides low-latency, composable, and distributed dataflow functionality. Applications are modeled as directed graphs, also known as Pipelines. DORA naturally offers a multi-agent, multi-process computation environment, as well as a foundation for modularity and composability. MoFA uses DORA as its executor to establish the Agent framework.
There are specific design patterns for building agents, and MoFA offers a series of implementations of these design patterns as templates for developers to use.
MoFA agents built on the Dora-RS framework essentially operate as Dora-RS Dataflows. Two agents can be combined via data connections, forming composite agents.
The self_refine Agent template mentioned above and the service agents and examples described below are all MoFA composite agents.
Agents often require certain services, such as retrieval-augmented generation (RAG), memory, external tool usage, and task planning/decomposition. MoFA follows the principle of "Everything is an Agent." It provides RAG agents, memory agents, planning agents, and action agents as MoFA agents. Developers can connect these pre-implemented service agents to their own agents to gain these services. MoFA also integrates third-party service agents for developers to use based on their needs.
- RAG
- Memory
- Action
- Planning
In general, you can use the Dora-RS command line to run agents:
-
Install the MoFA project package.
-
Execute the following command to start the agent process:
dora up && dora build your_agent_dataflow.yml && dora start your_agent_dataflow.yml --attach
-
If the first node of the agent dataflow is a dynamic node used for input/output in Dora, open another terminal and run
terminal-input. Enter the corresponding tasks to start the Agent process.