sfa-chat-graph

sfa-chat-graph is a retrieval-augmented generation system for RDF knowledge graphs that translates natural-language queries into SPARQL and executes them against the graph. It features a C#/.NET backend with a document-oriented database, REST and WebSocket endpoints for chat operations and real-time status updates, and an embedded secure Jupyter Notebook service for on-the-fly Python execution alongside an LLM API. The single-page frontend provides a conversational chat panel and an interactive force-directed graph visualization of each answer’s subgraph, complete with fragment downloads and detailed views.

It was built for the Swiss Federal Archives (SFA) by the University of Applied Sciences of the Grisons/Fachhochschule Graubünden (FHGR) in the context of the "SIARD zu Linked Data" project (2024-2025).

Configuration

Configuration of the System can be adjusted in the appsettings.json file, located in the sfa-chat-graph.Server folder. The following Settings can be made.

Path	Datatype	Default	Description
ConnectionStrings.Mongo	string	mongodb://mongodb:27017/sfa-chat-graph	Which Mongodb instance and database should be used by the backend
ConnectionStrings.Sparql	string	https://lindas.admin.ch/query	A stardog compatible endpoint for query access. Supports every Stardog database, no matter the contained graph
ConnectionStrings.Redis	string	localhost:6379	A optional redis database if caching should be done in redis and not in memory of the backend. If redis is added via docker this has to be adjusted to the containers hostname
Cache.Implementation	Enum CacheImplementation	InMemory	Which kind of caching should be used
Cache.DefaultExpiration	TimeSpan	00:30:00	How long chat-history should be retained in cache. Gets reset each time the user continues chatting
JupyterOptions.Endpoint	string	http://jupyter:8888	Which jupyter notebook server should be used for code execution
JupyterOptions.Token	string	-	Optional authentication token for the jupyter notebook server
JupyterOptions.SetupScript	string	pip install numpy pandas statsmodels matplotlib seaborn shapely geopandas[all]	bash script which should be run the first time the backend tries to execute code since the last restart
JupyterOptions.Kernel	string	python3	Which kernel should be used for code execution. If not specified the default kernel will be used. Should be any kernel which can execute python code
AiConfig.Implementation	string	Enum AiImplementation	OpenAI
AiConfig.Model	string	gpt-4.1	Which model should be used for inference. Important: The model needs to support tool calling.
AiConfig.ApiKey	string	-	API-Key for the AI Service
Aiconfig.ApiUrl	string	https://api.openai.com/v1/chat/completions	Endpoint which should be used for interaction with the chat inference api

CacheImplementation

Value	Description
InMemory	Caching should happen in memory of the application itself
Redis	Caching should happen in the supplied redis database
None	Caching should be disabled. Will lower memory usage in the cost of reponse time of the application

AiImplementation

Value	Description
OpenAI	Supports Openai compatible inference endpoints. Uses OpenAI per default

Deployment

The repo features a docker-compose.yml for easy setup. In the default configuration four containers are created. The outside ports can be adjusted to make accessible on other ports. If Express is not needed or wanted it can be disabled by commenting out the service in the compose file. The exposed port of the Backend is HTTP. If HTTPS is wanted using a reverse proxy like nginx is highly suggested. If a reverse proxy is not feasible Port 8081 of the backend can be exposed instead.

Container	Image	Required	Ports	Description
backend	sfa-chat-graph.Server/Dockerfile	yes	1221:8080	The backend of the system which also hosts the frontend website
jupyter	quay.io/jupyter/scipy-notebook	yes	-	The jupyter notebook server. It is used by the backend to execute python code in a secure sandbox. IMPORTANT: In Production the container should be configured in a way that it does not have outside internet access. This can be achieved by adding the container to the no-internet network. And using the image in jupyter/Dockerfile instead to install packages
mongo	mongo	yes	-	The Database the backend uses to store it's chat history. If gets to big, can be deleted, but will result in loss of all chat history
express	mongo-express	no	1222:8081	A webinterface to access the mongo db directly. Should only be enabled in dev environment

Environment Variables

The following environment variables can be bassed into the docker compose process.

Variable	Description
OPENAI_KEY	API-Key the backend should use to access OpenAI Services
EXPRESS_PW	The password for the mongo-express Webinterface