Tether

Prism AI is a chatbot designed to prevent hallucinations at the source rather than just filtering them out afterwards. Every question is first classified based on the type of risk involved, whether it's an incorrect fact, faulty reasoning, or a contradiction with something said earlier in the conversation. Based on this classification, a router that learns from previous outcomes decides which checking tier the question should go through. For factual queries, the response is broken down into smaller claims, each of which is verified against a Neo4j AuraDB knowledge graph. Calculations are checked step by step using Python, while follow-up questions are answered only using information that has already been verified in the conversation. Before generating the final response, a guard checks that everything is verified and rejects anything that isn't. If Prism AI cannot verify an answer, it simply says so instead of making things up. To make the system more efficient, verified reasoning is compressed using zlib and stored in SQLite so it can be reused for similar questions, reducing token usage. The project is built using a Meta model through the OpenAI-compatible API, with a FastAPI backend that streams responses using Server-Sent Events (SSE), a React + TypeScript + Vite frontend, and Docker for deployment.