logos-router-reasoning

Define a config.yaml listing each node with its engine (VLLM, Ollama, llama.cpp), model, host, port, and max thinking depth. Initialize the Router with that config, and it coordinates queries across nodes with consensus verification.

Enable Strict Write Discipline in the config with a consensus_threshold such as 0.95 and a peer_verification_count. Queries then require multiple nodes to agree before a response is written, and failures can trigger escalation or fallback.

Yes, multilingual routing supports 16 languages through a universal intermediate representation. Pass the input language and an output_language parameter, or supply mixed-language context documents with language set to auto.

Consensus fails when node agreement scores fall below the configured threshold, often due to model disagreement or ambiguous prompts. Enable debug mode to inspect failing nodes, disagreement reasons, and per-node verification scores.

The mesh supports local engines including VLLM, Ollama, and llama.cpp, each declared per node in the YAML configuration. Nodes can run on localhost for single-machine setups or on remote hosts for distributed GPU deployments.

Single-node inference fits latency-sensitive or low-stakes queries where consensus overhead is unjustified. The router supports forcing a specific node with consensus disabled as a fallback when mesh verification fails or times out.