We are helping financial companies with implementation of AI technology for fraud detection, compliance and document understanding. The industry is highly regulated and sensitive to mistakes and AI hallucinations. We have been asked several times to develop more reliable AI where the source of the data is only internal upstream systems and all returned results were explainable.
We have tested many techniques such as GraphRAG, chain of reasoning and agentic systems.
The most promising method is an automatic translation of natural language questions into multihop graph queries. This would help with hallucinations where the only source of the data became the updated knowledge graph and in the same time generated queries meant that each result left a signature of how and from where the information came and this solved the explainability issue.
We have tried to find open source or closed source tools that would give us acceptable results but it seems there are none generic enough and they suffer from brittleness of the generated queries.
We have decided to release an agentic system that we are developing as an open source this May. The amount of research and required expertise is high. We have gathered over 150 experts in the field who are interested in it so far. If you see that this is a worthy cause and you can help us spread the word it would be highly appreciated.
You can see bit more details at:
If you are interested please let me know.
Ladislav Urban
CEO of Dynocortex