RAG

Retrieval-Augmented Generation is a technique to improve LLM-generated responses by giving it additional knowledge. This typically involves gathering the information and injecting it into an agent’s system message for use by their LLM.

There are a number of ways to incorporate RAG into your AG2 workflow and agents:

• Use an AG2 reference agent, DocumentAgent (Coming Soon!)
• Add RAG capabilities to an agent
• Add RAG tools to an agent
• Incorporate context into an agent’s system message (manually and automatically)

​

1. DocumentAgent

Coming soon is our DocumentAgent, built specifically for RAG, and it will take the hassle out of loading, parsing, storing, and querying documents/web pages.

​

2. Add RAG capabilities to an agent

AG2 allows you to add capabilities to agents and an example of a capability is RAG using a graph database.

It only takes two steps to do this:

1. Create the capability
2. Add it to the agent

See the notebooks associated with the capabilities below for walk-throughs.

​

RAG Capability: Neo4j GraphRAG

Based on AG2’s base GraphRAG capability, this Neo4j GraphRAG capability allows the embedding and querying of information with a Neo4j graph database.

See the Using Neo4j’s graph database with AG2 agents for Q&A notebook.

​

RAG Capability: FalkorDB GraphRAG

Also based on AG2’s base GraphRAG capability, this capability uses a FalkorDB GraphRAG database.

See the Trip planning with a FalkorDB GraphRAG agent using a Swarm notebook.

​

3. Add RAG tools to an agent

Together with our upcoming DocumentAgent, will be the ability to add RAG tools directly to your agents. Watch this space!

​

4. Incorporating context into an Agent’s system message

ConversableAgent has a number of hooks that get run before an agent replies. You can utilise the update_agent_state hook to run a function that updates your agent’s system message with some context before it goes to the LLM.

Within the function use the ConversableAgent.update_system_message method to update the system message.

Let’s walk-through a simple example that puts a listing of the files in the current directory into an agent’s system message and asks an LLM to explain them.

We start with our imports, LLM configuration, and the system message template, which we’ll inject the file listing in to.

import os
from autogen import ConversableAgent, UserProxyAgent

config_list = {"model": "gpt-4o", "api_key": os.environ["OPENAI_API_KEY"]}

base_system_message = "You are a helpful agent, answering questions about the files in a directory:\n{filelisting}"

Here’s the function we’ll attach to the hook, it gets all files in the current directory and updates the associated agent’s system message accordingly.

def give_agent_file_listing(agent: ConversableAgent, messages: list[dict]) -> None:
    # Get the list of files in the current directory
    files = os.listdir()

    # Put them in a string
    files_str = "\n".join(files)

    # Use the system message template and update the agent's system message to include the file listing
    agent.update_system_message(base_system_message.format(filelisting=files_str))

Now we create the agent and attach the hook.

files_agent = ConversableAgent(
    name="files_agent",
    system_message="""You are a helpful agent, answering questions about the files in a directory.""",
    llm_config=config_list,
    )

files_agent.register_hook(
    hookable_method="update_agent_state",
    hook=give_agent_file_listing,
    )

Finally we create a human-in-the-loop agent and ask our files_agent about the files.

human = ConversableAgent(
    name="human",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=1,
    )

human.initiate_chat(
    recipient=files_agent,
    message="Tell me about the files in my directory.",
    max_turns=1,
    )

And we can see the LLM now knows about the files and directories in the current folder and is able to provide some information about them.

human (to files_agent):

Tell me about the files in my directory.

--------------------------------------------------------------------------------

>>>>>>>> USING AUTO REPLY...
files_agent (to human):

Certainly! Here's a brief explanation of each file and directory in your list:

1. **.ssh**: Contains configuration files and keys related to the SSH protocol, which are used to securely log into remote systems.

2. **.dotnet**: This directory is related to .NET and contains files and settings for .NET core and related development tools.

3. **.config**: A directory commonly used to store user-level configuration files for applications.

4. **.cache**: This directory is typically used to store cached data, which can improve the performance of the applications.

5. **ag2**: A directory for the open-source AgentOS, AG2 :)

These explanations are based on typical setups and usages; your specific use cases might vary. If you need more details on any item or it's a custom entry, checking the contents or configuration might be necessary for precise information.

--------------------------------------------------------------------------------