About the project

This project's goal is to have an agent that can generate Cypher queries to retrieve information from a graph database. The real chellenge will be the translation from natural language to Cypher and the explanation of the results. The technologies that will be used are: autogen, ollama, docker and python.

The main goal of the project is to have a system with two agents:

• Coder; generates queries and will explain the results in a natural language (English)
• Executor; runs the queries and return the result to the coder

Starting the project

Commands

To start a demo of the project you will need docker service running:

sudo systemctl start docker

To start the neo4j database you can launch the first command. If you want to run ollama with llama3.1 too, you can do so with the second command:

sudo docker-compose up
sudo docker-compose --profile local up

You need to launch this in the root directory of the project where docker-compose.yml is.

To run the python notebooks and files you firstly need to:

• Create a virtual environment

python -m venv .venv && \
source .venv/bin/activate

• install the dependencies

pip install -r requirements.txt

You can launch the notebooks after the start of all the components.

Important stuff

• If you have a computer that cannot run a llm model on a GPU you can try to run it with the CPU instead. Look in the docker-compose.yml to remove this part. If you want to use GPU instead you will need to load the correct drivers. You can find more details on the official docker documentation. If you want to serve a different model you can do it by changing the name of the model in the llm's healtcheck in the docker-compose. Remember to add your token/key if you are using a non local model.

• It is possible to install requirements.txt with poetry, you can do that with

poetry install $(cat requirements.txt) && \
poetry env use python

• You have to restart the notebook's kernel in order to make a new import effective. If it is not done it can generate an ImportError on the module you are attempting to import.

Project's structure

In the project you will find:

• tool_agent; implements an example on how to add tool usage to an agent
• simple_agent; implements an agent that should be able to generate cypher only with his previous knowledge
• code_executor_chat; implements a chat with a coder that can use his knowledge and some files to generate a query that is executed by the executor agent
• neo4j_tools.py; contains all the tools that operates on the neo4j database
• CypherExecutor.py; implements a custom code executor to run cypher queries

Project's state of the art

We have a code_generator_chat notebook that implements the type of conversation explained in About the project part. We will test if there is an improvement from a simple agent (has a good prompt and previous knowledge) to an agent that can be feed with manuals and other type of files to retrieve more information. For now both of them can work on task about:

• generate cypher queries
• question not concerning the database (as general purpose AI do)

Encountered prompt problems:

• if coder does not use a proper markdown format and if it does not separate two different queries an error will be trhown.
• if coder does not have the schema of the database and it tries to run the query it will probably fail due to some non existing relationships/nodes.
• if the rules for termination are given in the general rules in the prompt, the coder will terminate without passing the query to the executor
• Giving advices on how to write the query seems to generate without controlling the schema. Sometimes the llm will only generate a simple text using previous knowledge.

A future work will be adding more tools and agents.