UNDP Web Enabled AI Assistant

This repository contains a local, open source, privacy preserving, copyright respectful, interactive and transparent web search enabled AI assistant.

It often happens that the rollout of AI assistants in organizations conflicts with existing data and technology governance principles. This project aims to align principles we have come across with a light-weight AI assistant interface that runs in a terminal.

Setting things up

This code base is written in Python3. Make sure to have Python3.13 or above installed on your machine. It also requires that Ollama be installed on your machine, and that you have downloaded at least one open source Ollama compatible model.

• To download Ollama, visit: https://ollama.com/download.
• To browse and download different Ollama compatible LLMs, visit: https://ollama.com/library.

While it is configurable to use an online search engine, we strongly recommend you use an open-source local engine. Searxng is great for this! It is a privacy-respecting, hackable mastersearch engine—one that aggregates multiple different engines like Duck Duck Go, Google, etc.

• To run Searxng locally, we recommend you download and run the Docker image here: https://github.com/searxng/searxng-docker.
• Make sure to configure the searxng/settings.yml as follows:

search:
	...
	formats:
		- html
		- ...
		- json

The AI agents use the json format. If you do not use Searxng, or do not activate the json format, you will need to edit the ./process/_03_perform_search.py file to return a json object with the appropriate formatting.

This of requires that you have Docker installed on your machine. Once Searxng is installed, run it via Docker. This should start a localhost server on port 8080.

When these requirements are met, create a local python environment by running:

python -m venv /path/to/virtual/environment

Then install all the requirements:

pip install -r /path/to/requirements.txt

That is all! You should be good to go. To start the assistant, run:

python main.py