This is EcoToxFred, a Neo4j-backed Chatbot with whom you can discuss environmental monitoring data collected in a large knowledge graph and stored in a Neo4j Graph Database.
EcoToxFred is a prototype that demonstrates how domain-specific knowledge can be combined with LLMs to facilitate the interaction between users and (domain-specific) knowledge.
EcoToxFred can answer questions about chemicals and their measured concentrations in European surface waters like rivers and lakes.
EcoToxFred is a friendly and intuitive chatbot designed to help researchers, stakeholders, and users explore and analyze chemical concentrations and toxicity data in European surface waters.
By leveraging advanced language models and a graph database, EcoToxFred makes complex environmental toxicology data
accessible and understandable for everyone, regardless of their technical background.
In its currently trained state, it can
- provide general information about rivers, lakes, or substances
- provide sampling sites where individual (sets of) chemicals have been measured and detected at which concentrations
- plot measured concentrations across time and/or sampling sites
- find all substances measured at a certain sampling site (identified by country, water body, or coordinate range)
- identify the most frequent drivers in regions (countries, water bodies, Europe)
- extract summarized risk for regions or species
Upcoming skills are
- plot requested information on a map of the selected region
- provide subgraphs of graph database results, ready to be used in Cytoscape or else
To run the application, you must install the libraries listed in requirements.txt.
pip install -r requirements.txt
Then run the streamlit run command to start the app on localhost:8501.
streamlit run bot.py
If you prefer to use Docker, you just can run the app including the Neo4j-Database with:
docker compose up --renew-anon-volumes --buildYou may experience that EcoToxFred does not provide appropriate responses. Please provide your query and your expected response, and we will improve LLM prompts, provide more examples for zero-shot and few-shot learning, and if necessary even make improvements or adjustments to our Neo4j database (e.g., by integrating additional knowledge).
Jana Schor - [email protected]