# EC-RAG Agent Guide

This guide explains how to enable, configure, and use Agent capabilities within EdgeCraftRAG (EC-RAG). It covers the creation and management of different agent types to enhance retrieval and generation workflows.(Currently, EC-RAG Agent is an experimental option.)

## Overview

EC-RAG introduces autonomous agents that can perform complex reasoning, planning to answer queries more effectively than standard RAG pipelines.

EC-RAG currently supports two types of agents:

- **Simple Agent**: A lightweight agent designed for straightforward tasks and direct query answering.
- **Deep Search Agent**: An advanced agent capable of multi-step planning, deep research, and reasoning.

## Creating an Agent

### Prerequisites

To use agents in EC-RAG, please ensure the following prerequisites are met:

1. Ensure the standard EC-RAG services are running and the UI is accessible.
2. Agents require vLLM serving, please ensure pipelines utilizing vLLM serving are available.

### Creating a Simple Agent

To create a Simple Agent, you need to switch to `Agents` tab in the `Settings` page and click `Create Agent` button.
![alt text](../assets/img/agent_tab.png)

Then in the `Create Agent` window, set `Agent Type` as `Simple`.
![alt text](../assets/img/agent_create_simple.png)

After selecting `Simple`, pipeline selection and Agent configuration will be available. Please select the pipeline with vLLM serving (generator type `FreeChat`). For `Agent Configs`, you may specify `Max Retrievals` as needed.
![alt text](../assets/img/agent_config_simple.png)

### Creating a Deep Search Agent

To create a Deep Search Agent, switch to `Agents` tab and click `Create Agent` button. In `Create Agent` window, set `Agent Type` as `Deep Search`.
![alt text](../assets/img/agent_create_deep-search.png)

Similarly, please select the pipeline with vLLM serving (generator type `FreeChat`) in `Pipeline Name`. In `Agent Configs` session, 5 parameters could be customized: `Retrieve Top K`, `Rerank Top K`, `Mece Retrieval`, `Max Retrievals` and `Max Plan Steps`. If `Mece Retrieval` is enabled, agent will not retrieve chunks already retrieved in each step.
![alt text](../assets/img/agent_config_deep-search.png)

## Using Agents in RAG Pipelines

When agents were created successfully, they will be listed under `Agents` tab.
![alt text](../assets/img/agent_tab_with_agents.png)

Then we can switch to `Chat` page to use the agents we created. They will be listed under `Agents` section of `Chat` page.
![alt text](../assets/img/agent_chat.png)

### Simple Agent

Select the Simple agent we created, `Think` function will be locked in dialog window as `Think` is mandatory for agents. While Simple agent is working, it will decide if more information is needed, and generate new query if so.
![alt text](../assets/img/agent_simple_working.png)

### DeepSearch Agent

Select DeepSearch agent from the `Agents` section. If we're going to switch between different agents, previous will be close.
![alt text](../assets/img/agent_switching.png)

Agent could also be deactivated by single-click on the activated agent.
![alt text](../assets/img/agent_deactivate.png)

While DeepSearch agent is working, it will make plans on its own and execute the plan step by step.
![alt text](../assets/img/agent_deepsearch_plan.png)

When all planned steps completed, DeepSearch agent will generate the final answer.
![alt text](../assets/img/agent_deepsearch_answer.png)