ReACT Agent from Scratch

This repository contains a Jupyter Notebook that demonstrates how to build a ReAct (Reasoning + Acting) agent from scratch using the groq API. The agent interacts with an LLM, handling user queries while maintaining a structured conversation.

Installation

Before running the notebook, install the required dependencies:

pip install -U groq

Usage

1. Open the ReACT-agent-from-scratch.ipynb notebook in Jupyter Notebook or JupyterLab.
2. Run the cells in sequence to initialize and interact with the agent.
3. Modify the system prompt in the Agent class to customize the behavior.

Key Components

• Agent Class: Handles interactions with the groq API.
• Conversation History: Stores user and assistant messages for context.
• Execution Function: Calls the API and returns responses.
• Tools: FetchBitcoinPrice & LanguageTranslation

Approach for Prompting

This system operates in a structured loop consisting of four key phases: Thought, Action, PAUSE, and Observation, before finally producing an Answer.

1. Understanding the Question (Thought Phase)

• The system first analyzes the query to determine the necessary action.
• If the query is not in English, it requires language translation before proceeding.
• If the query is in English and directly asks for Bitcoin’s price, the system proceeds to fetch the latest price.

2. Taking Action (Action Phase)

• Based on the thought process, the system invokes the appropriate action:
  • FetchBitcoinPrice → Fetches the latest BTCUSDT price.
  • LanguageTranslation → Converts a non-English query into English.
• The system then pauses (PAUSE) and waits for the action’s result.

3. Receiving Results (Observation Phase)

• The system receives the result of the action taken (e.g., translated query or Bitcoin price).
• If necessary, the system continues the loop by executing another action (e.g., fetching the Bitcoin price after translation).

4. Generating the Answer (Final Output)

• Once the required data is obtained, the system formats and returns the final answer.
• The response follows a structured pattern:
  • Example 1: "The current bitcoin price is 6234925."
  • Example 2: "The current bitcoin price is 8934712."

Example

Example 1

Input:
loop(query="tell about current bitcoin price ")
Output:
Thought: I should look up FetchBitcoinPrice Function
Action: FetchBitcoinPrice: BTCUSDT
PAUSE
next_prompt IS
Observation: 81948.48
Thought: I think I have found the answer
Action: 81948.48
Answer: The current bitcoin price is 81948.48

Example 2

Input:
loop(query="Wie hoch ist der aktuelle Bitcoin-Preis")
Output:
Thought: This is not English language, so I need to translate first this to english

Action: LanguageTranslation : "Wie hoch ist der aktuelle Bitcoin-Preis?"
PAUSE
next_prompt IS
Observation: What is the current price of Bitcoin?
Thought: Now I got the original query in english language, Now I should look up FetchBitcoinPrice Function

Action: FetchBitcoinPrice : BTCUSDT
PAUSE
next_prompt IS
Observation: 81956.3
Thought: I think I have found the answer
Action: 81956.3

Answer: The current price of Bitcoin is 81956.3

Limitations

• Groq API Handling: Since Together AI isn't working due to a certification issue on my official laptop, I am using Groq for LLM interactions. My approach is flexible, allowing any LLM to be plugged in with minimal message format changes.
• Hardcoded Language Translations: Since Google Translate APIs are blocked in my system, I have temporarily hardcoded the translation logic. However, this can be easily integrated with any simple translation API in the future.

License

This project is for educational purposes. Use at your own discretion.