How to Create a New Assistant

This guide describes the complete flow for creating an assistant on the Admin Platform, whether it uses knowledge augmented retrieval (RAG) or is purely conversational.

Overview

When creating an assistant you define:

1. Basic information (name, description, provider/model)
2. Prompts and parameters for the system, user, and response
3. Mode of operation: with or without RAG
4. Knowledge collections (when RAG is enabled)
5. Optional MCP tools and workflows

Prerequisites

• Have at least one configured provider/model
• Have ready Knowledge Collections if you wish to use RAG
• Access permission at CURATOR level or higher to the project where the assistant will be created

Step 1: Access the Creation Page

1. Log in to the admin platform.
2. In the main sidebar menu, select Projects and choose the desired project.
3. Click on Assistants in the project sidebar menu.
4. Press the + New AI button in the upper right corner.

Step 2: Select AI

By clicking on + New AI, you will be directed to the creation flow. The first step is to select the type of AI you want to create.

1. On the "New AI" screen, you will see the Assistant option available
2. Click on the "Assistant" card to create a custom AI assistant
3. Click Next to continue

Step 3: Assistant Details

In the second step, you will configure the basic information of the assistant:

Fill in the following fields:

• Name (required): A descriptive name to identify the assistant
• Type of assistant (required): Choose between:
  • With RAG: For assistants that use knowledge augmented retrieval
  • Without RAG: For conversational assistants without access to knowledge collections
• Description: An optional description of the assistant's purpose and functionalities

After filling in, click Next to continue.

Step 4: Select the Model

In the third step, you will select the AI model that will be used by the assistant.

1. Review the available models in the list
2. Click on the desired model (e.g., "GPT-4o")
3. Select the specific version of the model in the "Select version" dropdown
4. You can click on the information icon (ⓘ) to see details of the model, including:
   • Token limits (input and output)
   • Costs per token
   • Capabilities (Function Calling, Vision, etc.)
   • Available versions

After selecting the model and version, click Next to continue.

Step 5: Choose the Mode (Without RAG vs. With RAG)

Assistant Without RAG

• Ideal for flows where all context is in the prompt or in MCP tools.
• Just confirm the prompts and parameters; no knowledge source is loaded.

Assistant With RAG

1. If you selected "With RAG" in the previous step, this step will allow you to select the knowledge collections.

2. Select Knowledge Collections: Click on the cards of the collections you want to use. The selected collections will appear highlighted in blue.
3. You can select multiple collections to combine different sources of knowledge.
4. Configure the number of documents retrieved. This is the total number of excerpts that will be provided to the AI to generate the response. Higher numbers (>10) provide more context but can confuse the AI. The optimal value depends on the application and should be experimented with.
5. Adjust specific fields like context size, search temperature (controls the creativity of responses), and token limits (maximum amount of text that can be processed).

To create collections or sources, refer to the Introduction to Knowledge Collections and the Introduction to Knowledge Sources.

Step 6: Select MCP Servers

You can enable MCP integrations to allow the assistant to perform external actions (query APIs, trigger automations, etc.).

To select MCP servers, you must first install them in the MCP Servers section of the project. Access this section through the project sidebar menu.

On the MCP Servers screen, you can:

• View Installed servers
• Search and install servers from Docker Hub
• Search and install servers from GitHub

To install a server from Docker Hub:

1. Click on the Docker Hub tab
2. Use the search bar to find the desired server
3. Click on the server card to see the details
4. Review the server information, including:
   • Server description
   • Author and Docker image
   • Required command
   • Required environment variables
5. Configure the necessary environment variables (e.g., BRAVE_API_KEY for the Brave Search server)
6. Click Install

After installing the MCP servers, you will be able to select them in the assistant creation step:

1. Configure the Maximum Tool Iterations: Define how many times the tools can be executed in sequence (e.g., 5). Higher values give more freedom to the AI but may cause it to enter an error loop for more complex issues.
2. Select the MCP servers: Click on the MCP servers you wish to enable (maximum of 5 servers)
3. The selected servers will appear highlighted in blue
4. Examples of available servers: "bible," "youtube_transcript," "playwright," etc.

After configuring, click Next to continue.

Step 7: Workflow Details

In the final step, you will configure the workflow details and prompts of the assistant.

Configure the following options:

1. Number of Retrieved Fragments: Define how many document fragments will be retrieved for context (e.g., 5)
2. Retrieve full documents: Check this option if you want to retrieve full documents instead of just fragments
3. Edit your prompts:
   • Rephrase Question: Prompt that reformulates the user's question independently of the conversation context
   • Respond to User: Prompt that guides how the assistant should respond based on the provided context

After reviewing all configurations, click Save to create the assistant.

Step 8: Confirmation

After clicking Save, you will see a success notification confirming that the assistant has been created.

The assistant will be immediately available in the assistants list and can already be used in conversations.

Next Steps

• Manage Assistants
• Configure Knowledge Collections
• Connect Knowledge Sources