Compatibility matrix

Component	Supported
Python	3.11, 3.12, 3.13 (CI matrix on every PR)
OS	Linux / macOS (Windows best-effort)
Default model	mock:echo offline; providers via extras
Docs site	MkDocs Material → GitHub Pages (Docs workflow)

Optional extras

Extra	Purpose
serve	FastAPI gateway / UI
redis	Redis cache + queue worker
eval	sacreBLEU + sentence-transformers cross-encoder
ocr	PDF OCR (pypdfium2 + pytesseract)
pgvector / neo4j / peft / training / torch	Store / train backends

Live integrations

Opt-in probes live under tests/live/ and are skipped unless env vars are set:

• AIRE_LIVE_REDIS
• AIRE_LIVE_PGVECTOR
• AIRE_LIVE_NEO4J
• AIRE_LIVE_QDRANT
• AIRE_LIVE_CHROMA
• AIRE_LIVE_WEAVIATE
• AIRE_LIVE_MILVUS

Run: pytest -m live.

Optional CPU backend checks run with OMP_NUM_THREADS=1, OPENBLAS_NUM_THREADS=1 and KERAS_BACKEND=torch; mixing unbounded native runtimes in one process is not a tested performance configuration. GPU quantization needs compatible bitsandbytes hardware and is not verified by CPU or dry-run tests. The capability matrix records concrete evidence and limits.