Skip to content

README.md

Latest commit

 

History

History
84 lines (53 loc) · 2.17 KB

README.md

File metadata and controls

84 lines (53 loc) · 2.17 KB
Raw

✅ Prerequisites

  • Python 3.8 or later
  • Windows (commands provided are for Windows; minor adjustments may be needed for macOS/Linux)

🚀 Setup Instructions

🔑 Step 1: Get OpenRouter API Key

  1. Go to OpenRouter API Keys
  2. If you don't have an account, create one first.
  3. Click Create API Key.
  4. Copy and save your key securely — you can only view it once.

⚙️ Step 2: Set Environment Variable

In VSCode terminal, run:

$env:OPENROUTER_API_KEY = "insert_your_openrouter_api_key_here"

📦 Step 3: Install Dependencies

Install required libraries using:

pip install -r requirements.txt

📄 Information for Users

1. 1_multi_agent_deepseek.py

This program initializes 4 instances of the DeepSeek LLM as 4 different agents:

  1. Assistant
  2. Planner
  3. Coder
  4. Critic

When the user gives a task as a prompt, these 4 agents work collaboratively to solve it. By default, 5 rounds of conversation are set up where each agent gets the complete history of messages to work on before giving its response.

2. 2_multi_agent_with_summarizer.py

This is the same as 1_multi_agent_deepseek.py, except it has an additional summarizer agent. The summarizer compiles the conversation between the first 4 agents as the 5th step, then passes the summary into the next round. This helps ensure the LLM input word limit is not exceeded.

3. 3_Evaluate_multiple_models.py

This script evaluates 4 models:

  1. DeepSeek V3 0324 (free)
  2. OpenAI: gpt-oss-20b (free)
  3. Z.AI: GLM 4.5 Air (free)
  4. MoonshotAI: Kimi K2 (free)

They are tested on 3 hardcoded tasks for:

  • Accuracy — compared against expected outputs
  • Efficiency — shorter responses are better
  • Consistency — each model responds 3 times to the same prompt; results are compared for similarity
  • Robustness — each model receives an original and slightly modified prompt; responses are compared

The hardcoded tasks and expected results were generated using OpenAI GPT-5.

Future Work: Automate the step of generating tasks and expected results using an LLM.