Home

Open-ASR Model Explorer: AgentWiki

Welcome to the active knowledge base for the Zero-Overhead Data Path and hybrid inference routing architecture.

Directories

• 00_Raw_Context/ - Drop raw documentation here for the agent to ingest.
• Concepts/ - Atomic technical concepts (e.g., LocalAgreement-2).
• Architecture/ - System design and routing patterns.
• Logs/ - The agent's changelog.

Concepts

Page	Summary
LocalAgreement-2	Client-side streaming stabilization: holds back text until two consecutive hypothesis passes agree.
Chunked_Prefill	vLLM scheduler setting that prevents long audio prefills from starving concurrent decode streams.
WebGPU_Cache_API	Browser Cache API strategy for persisting ONNX model weights across page refreshes.

Architecture

Page	Summary
Inference_Router_Strategy	Server-side routing of Cohere and Qwen to FastAPI/vLLM, including audio normalization and chunked prefill config.
Deployment_Topology	Multi-container Docker Compose deployment for backend and frontend, including Nginx SSE proxy constraints.