RAG system with Agent is a good syatem for single step question answering. It means a simple Agent can dynamically choose between tools based on current state and perform a single pass operation through pipeline tools to generate response. A complex task may require breaking down a instruction into subtaska and perfrom sequentially. When performing sub taks sequentially we also need a buffer memory to keep track of the works. This powers is incorporated into agent with a reasonong loop in Llama Index.
A simple Agentic-RAG is good for 1-to-1 question answering. However a complex task may requiire sequence of subtasks to be executed to arrive at the final answer. This is where reasoning loop comes into play. In a reasoning loop the agent is able to perform subtaks sequentially until final response.
A agent consist of Agent Runner and Agent Worker. These two componnt interact in a cyclic manner until final response. Agent worker perform the actual work of tool selection through reasoning via chain of thoughts and deciding next step or formulating the final output. Whereas Agent Runner is the orechastator responsible for mainintaining task state and memory buffer. It works more like a scheduler and dispatcher.
The Image below shows the Agent Reasoning Loop with Agent Component and their role.
image refernec:DeepLearning.AI
On a broader term the Reasoning Loop perform tasks in a logical sequence. Well what is the sequence. How every component is interacting?
To understand the workflow we need to undestand chain of thoughts. To break a complex task into subtask LlamaIndex Agent makes use of chain of thought to guide the LLM into deciding the next action. These chain of thought are introduced by template with some fixed set of prompts and some dymanic context relevant to the current step.
The diagram below shows the sequence of events in an AGENT REASONING LOOP
- Agentic-RAG: Llama Index
- App: Gradio
- LLM: Llama3 8B
- Embedding: nomic-embed-text
- Local LLM: Ollama
- Containerization: Docker
- agent-reasoning-loop.ipynb: Code for implementing an RAG Agent with Llama Index.
- app.py: Code Gradio application. Tools creation functions are present in
utils.pyand AGENT creation funtions are present inget_agent.py
We need a LINUX system with atleast 8GB RAM.
- Install libraries with
make install - To Run the Application execute
docker compose upIf you want to run the notebooks - Download Ollama Docker Image and start Ollama server with
ollama_dockeron a new CLI as the server will block the CLI. - Now
agent-reasoning-loop.ipynbNotebook can be executed. If previous images or container exist with same name it may create conflict. Delete image/conatiners if name conflict occurs.
- Thanks to DeepLearning.AI and LlamaIndex for the wonderful course
- Thanks to
Metafor open source Llama3