Thanks for your input. It's deterministic because no embeddings. TF-IDF is used because it's not vector based and doesn't rely on vibes. Still figuring out how to make it better. If you can help suggest, that would be great.
I built QWED – a verification layer that sits between your LLM and production.
The idea: Don't fix hallucinations, verify them. If AI output can't be mathematically proven, it doesn't ship.
11 specialized engines:
- Math (SymPy) – verify calculations
- Logic (Z3 SMT) – formal proofs
- SQL (SQLGlot) – detect injection/dangerous queries
- Code (AST) – security analysis + taint tracking
- Facts (KB) – entity verification without LLM
Works with ANY LLM – OpenAI, Claude, Gemini, or local models via Ollama ($0).
Model-agnostic: Your LLM choice, our verification.
Happy to answer questions about deterministic AI verification!