Back to projects

Polaris

Active

Multi-model AI orchestration engine with self-healing RAG, phase-based development loops, and multi-model delegation

PythonLanceDBVoyage AItree-sitterMCPReact

Architecture

Autonomous dev loops with phase orchestration, QA gates, and drift detection. Multi-model routing delegates tasks across Claude, GPT-4o, Gemini, and Kimi K2 based on capability matching.

Data Pipeline

AST-aware chunking via tree-sitter, voyage-code-3 1024-dim embeddings, LanceDB vector storage, Voyage Rerank 2.5

Model Integrations

ClaudeGPT-4oGeminiKimi K2

Novel Approaches

  • •Self-healing RAG with drift detection and auto-reindex
  • •AST-aware code chunking (functions/classes as semantic units)
  • •Multi-model delegation via polaris-ask CLI with parallel execution
  • •Two-pass QA: automated (pytest/mypy/ruff) + adversarial verification agent
  • •Persistent state via Polaris.md visible in PRs

Key Files

polaris/core/index.py
polaris/core/orchestrator.py
polaris/chunker/bridge.py
polaris/routing/router.py
polaris/mcp/server.py