I'm blown away by what we have today. Take https://huggingface.co/Qwen/Qwen3-Embedding-0.6B which is just brilliant at taking output from RAPTOR and the datasets we create from our /understand (attack path, src, sink, dataflow, threat model etc) and then doing careful extraction, not crystal-ball nonsense. It’s capable enough to understand technical security language and follow a structured schema, but small enough to run properly on an 8GB M1 Mac mini. I run it with temperature 0, an 8,192-token context and a tightly defined output schema, num_predict = 1400 which bounds the output. The model has enough room for the structured labels but cannot wander into an enormous response or hallucinate as much in my experience. Then, I use qwen3-embedding:0.6b to spot semantically related labels Honestly this is such an exciting period for us, well at least me finding bugs 'n shit. on the dinosaur front, we set about making a bloody solid sandbox and control/telemetry plane for RAPTOR: https://github.com/gadievron/raptor/issues/889 oh and it works really well from stopping it from eating your face.