ADR-0005: Python tech stack for MVP

Date: 2026-06-15
Status: accepted

Context

daari needs a primary implementation language (OD-2). Options evaluated in docs/discovery/03-approach-options.md. Plan review issue #9.

Decision

Python 3.12 for Phase A through Phase B implementation.

Rationale

Factor	Python	Go
MVP velocity	✅ Fast iteration	Slower
FastAPI OpenAI-compat	✅ Mature	Build more
Embeddings / eval (L1, routing tests)	✅ Rich ecosystem	Weaker
Ollama clients	✅ httpx, ollama lib	Available
Daemon packaging	⚠️ venv/pyinstaller	✅ Single binary
Solo builder	✅ One person, fewer files	More boilerplate

Trade-off accepted: Packaging complexity over raw startup latency. daari is localhost-bound; 50ms vs 5ms startup is irrelevant.

Stack

Python 3.12
├── FastAPI + uvicorn     # OpenAI-compat gateway
├── typer                 # CLI (serve, stats, setup, doctor)
├── httpx                 # Ollama + frontier HTTP
├── pydantic + pydantic-settings  # config
├── diskcache or sqlite3  # L0 cache
└── pytest                # routing eval + unit tests

External dependency: Ollama (user-installed, not bundled).

Consequences

• ADR revisit if p99 latency or memory becomes problem at scale
• Optional Go rewrite of gateway only in v2 — router logic stays testable in Python
• CI: Python 3.12, pytest, optional Ollama integration job

Not chosen (for now)

Option	Reason deferred
Go	Better daemon story but slower ML/routing iteration
Rust	MVP velocity too slow
TypeScript	Weaker local embedding/eval path