# Experiments with LLMs

This repository contains experiments with LLMs.

IMPORTANT: Each `python` command can be replaced with `uv run python` if you are using uv. The `pyproject.toml` file is already configured to create a virtual environment with the correct dependencies.


- [Experiments with LLMs](#experiments-with-llms)
  - [Requirements](#requirements)
  - [Setup](#setup)
    - [Env file](#env-file)
    - [Install dependencies](#install-dependencies)
  - [Experiments List](#experiments-list)
    - [Artificial COT](#artificial-cot)
      - [How to use](#how-to-use)
      - [How it works](#how-it-works)
      - [Customizing the script](#customizing-the-script)
    - [Self Conversational AI](#self-conversational-ai)
      - [How to use](#how-to-use-1)
      - [How it works](#how-it-works-1)
      - [Customizing the script](#customizing-the-script-1)


## Requirements

- An `ollama` server running somewhere.

## Setup

### Env file

Copy the `.env.example` file to `.env` and set the variables to your liking.
Be sure you can reach the ollama server from the machine running the scripts.

### Install dependencies

NOTE: This step is not required if you are using uv.

```bash
pip install -r requirements.txt
```

## Experiments List

### Artificial COT

This is a simple implementation of the an artificial COT method.

The idea is to have an assistant that can generate a Chain of Thought (COT) for a given problem and use it to solve the problem in a second answer.

The approach can use any LLM to generate the COT and the second answer, even different models for each step.

#### How to use

```bash
python src/artificial_cot.py
```

To modify the prompt, you can edit the `prompt` variable at the top of the script.

#### How it works

The `artificial_cot.py` file instantiates two `OllamaConnector` instances, one for the COT and one for the response.

The system prompts are overriden to customize the behavior of the COT and the response.

The script will use the model specified in the `.env` file as the model for the COT.

#### Customizing the script

You can use `cot_ollama.set_model()` and `response_ollama.set_model()` to change the model for the COT and the response.

You can also change the system prompts to customize the behavior of the COT and the response.

### Self Conversational AI

This experiment aims to create a self conversational AI.

The idea is to have an AI that can converse with itself or with another LLM model.

#### How to use

```bash
python src/self_conversational_ai.py
```

#### How it works

The `self_conversational_ai.py` file instantiates two `OllamaConnector` instances, with the same system prompt and model.

The script will then enter a loop where it will:

1. Generate an initial greeting to kickstart the conversation.
2. Feed the greeting to the other LLM instance.
3. Feed the response to the first LLM instance.
4. Repeat the process until the conversation ends (CTRL+C)

The script will override the system prompt using the `self_system_prompt` variable at the top of the script.

#### Customizing the script

You can change the `greeting` variable at the top of the script to customize the initial greeting.

You can also change the models by using `first_ollama.set_model()` and `second_ollama.set_model()` to change the model for the first and second instances.

Same goes for the system prompts.