Model Support
The intent of TrustGraph is to be model agnostic. While there is support for common LLMs through AWS Bedrock, Anthropic, AzureAI, Cohere, Ollama, OpenAI, and VertexAI, the true power of TrustGraph is the ability get "big model performance" with the latency and cost efficiencies of open source Small Language Models (SLMs). TrustGraph has been fully tested with the following SLMs:
Gemma-2-9BPhi-3-SmallMixtral8x7BDeepSeek-V2-16BAya:8BLlama3.1:8BQwen2:7B
While Llama3.1 scores highly in many metrics, testing with TrustGraph for the purpose of naive extraction has yielded unexpected results. Llama3.1 tends to narrowly scope the concept of entities tigher than other models. In other words, other models like Gemma-2, Phi-3, Mixtral8x7B, DeepSeekV2, and Aya extract a wider range of conceptual entities. Llama3.1 tends to extract entities that relate to people, whether they are inviduals, groups of people, or organizations of people. Interestingly, identifying individuals is a weakness of other models. For these reasons, it is very difficult to definitively say which model is "best". Instead, TrustGraph supports as many models as possible to allow for finding the best model for a given use case.