Every infrastructure and AI package in QuadKit separates caller code from backend drivers: your services inject a protocol from quadkit-contracts, application.yaml selects the backend, and optional extras pull only the SDKs you run in production.
Availability: quadkit-ai-llm is published at 0.0.50, alongside the SQL and cache packages.
uv add "quadkit-ai-llm[anthropic,openai]"
uv add "quadkit-sql[postgres]"
uv add "quadkit-cache[redis]"
Provider Extra Best for Key trade-off anthropic[anthropic]Tool-heavy agents, long context, coding & structured reasoning API cost on frontier tiers openai[openai]Broad model catalog, structured outputs, multimodal workloads Cloud API dependency ollama[ollama]Local dev, offline CI, strict data residency Requires local GPU/CPU capacity groq[groq]Ultra-low-latency token streaming on open-weight models Context window & rate caps vary by tier mistral / cohere[mistral], [cohere]Multilingual generation, EU residency, enterprise rerank/chat Smaller ecosystem than OpenAI/Anthropic openrouter(built-in HTTP) Single key across dozens of hosted models & fallback chains Additional gateway hop
Backend Extra Best for Key trade-off sqlite[sqlite]Local development, single-node apps, fast integration tests Single writer; limited ALTER TABLE postgres[postgres]Production default — transactional workloads, JSONB, concurrency Separate database service to operate mysql[mysql]Existing MySQL/MariaDB fleets and read-replica topologies Fewer JSON/index capabilities than PostgreSQL
Backend Extra Best for Key trade-off memory(built-in) Local dev, tests, single-process services Process-local; cleared on restart redis[redis]Production caching, stampede locks, shared TTL state Requires Redis/Valkey memcached[memcached]High-throughput volatile key-value caching No persistence or rich data structures
Backend Extra Best for Key trade-off local(built-in) Development, single-host deployments, test fixtures Host-bound disk; no built-in CDN s3 / S3-compatible (r2, MinIO)[aws]General object storage, presigned uploads/downloads Egress fees on AWS S3 (zero egress on Cloudflare R2) gcs[gcp]GCP-native workloads Tied to GCP IAM & buckets azure[azure]Azure-native workloads Tied to Azure Blob Storage
Backend Package / Extra Best for Key trade-off memoryBuilt into both In-process domain events, tests, single-worker jobs Lost on process exit sqlite / postgresquadkit-events[sqlite,postgres]Durable event store, transactional outbox, projections Throughput bounded by SQL write rate redisquadkit-tasks[redis]Low-latency background job queues & scheduled tasks At-most-once unless paired with retries/DLQ rabbitmqquadkit-events[rabbitmq], quadkit-tasks[rabbitmq]Durable routing, acknowledgements, dead-letter exchanges Broker operational overhead kafkaquadkit-events[kafka]High-volume event streaming, consumer groups, replay Heaviest operational footprint
For multi-step business processes with retries, human approvals, or compensating rollbacks across services, pair events/tasks with quadkit-workflow
Channel Extra Typical use console / memory(built-in) Local development, deterministic test assertions, in-app inbox smtp / sendgrid(built-in SMTP) , [sendgrid]Transactional email, digests, verification links twilio[twilio]SMS alerts and verification codes apns / web-push[apns], [web-push]Mobile and browser push notifications slack / whatsapp[slack], [whatsapp]Operational alerts and customer messaging
Capability Extra When to enable Health checks (/health, /ready) (built-in) Every service — aggregates provider health_check() status automatically Host & process gauges [system]CPU, memory, and file-descriptor metrics via psutil Prometheus /metrics [prometheus]Scraped Kubernetes / VM monitoring stacks OpenTelemetry OTLP [otel]Distributed traces and metrics exported to Grafana Tempo, Jaeger, Honeycomb, or Datadog
