logos-distributed-reasoning-router

Clone the repository, install dependencies with pip, and create a YAML config defining your mesh nodes with engines like VLLM or Ollama. Then initialize the Router class with that config and submit queries with a chosen reasoning depth.

Each reasoning step is proposed by a node, broadcast to peers for verification, and committed only when confidence exceeds the consensus threshold. Failed consensus triggers escalation to a higher reasoning depth with an arbiter node.

Yes, the mesh configuration supports local inference engines including VLLM, Ollama, and llama.cpp. Each node declares its engine, model, host, port, and maximum thinking depth in the YAML config.

Yes, it supports 16 languages with universal semantic representation. You can set input and output languages independently per query, for example asking in Chinese and receiving the response in English.

Low consensus usually stems from unhealthy nodes, too few verification peers, or an overly strict threshold. Check node health scores, increase verification_peer_count, or temporarily lower the consensus threshold while inspecting failed audit trail steps.

Profile query timing to separate consensus overhead from processing and network latency. Enable semantic caching, use lower fixed depth for simple queries, and rebalance the mesh to distribute load across nodes.