| daari |
Open-source local execution router (Telugu: path, way) |
| Tier |
A level in the routing stack (L0–L6, Lt) |
| Path |
The tier chosen for a specific request |
| L0 |
Exact cache — identical prompt → instant hit |
| L1 |
Semantic cache — similar meaning → reused response |
| L2 |
Rules — deterministic transforms, no model |
| Lt |
Tool-native — IDE/CLI execution without AI |
| L3 |
Small local model (SLM) |
| L4 |
Medium local model |
| L5 |
Large local model |
| L6 |
Frontier cloud API (OpenAI/Anthropic) — last resort |
| $0 tier |
L0, L1, L2, or Lt — zero marginal inference cost |
| CCS |
Command context store — reuse command output across turns |
| L2-live |
Live factual rule profile (weather, search) |
| ProviderRegistry |
Plugin system for all backends (ground level Phase A) |
| IntegrationProvider |
Protocol — MCP, Sourcegraph, skills, open APIs, Google implement this |
| PolicyEngine |
Lt deny/ask/allow before execution — ADR-0012 |
| Lt-fetch |
External API/search fetch via registered providers (subset of Lt) |
| Router |
Classifies requests and selects tier |
| Executor |
Runs the work at a tier (tool, Ollama, frontier) |
| Frontier |
Cloud LLM providers (OpenAI, Anthropic) |
| Local-first |
Prefer on-machine tiers before cloud |
| Tracer bullet MVP |
Smallest shippable proof of cost/latency wins |
| L0-org / L1-org / CCS-org |
Enterprise shared cache tiers (Phase E2) — after local miss |
| Org control plane |
Self-hosted fleet policy, org cache, learning (Phase E1–E3) |