// the find
instill-ai/instill-core
🔮 Instill Core is a full-stack AI infrastructure tool for data, model and pipeline orchestration, designed to streamline every aspect of building versatile AI-first applications
Instill Core is a self-hosted AI pipeline platform — think n8n but purpose-built for ML workloads. It wraps ETL, RAG, model serving, and workflow orchestration into a single Docker Compose stack. Aimed at teams who want to run AI pipelines on their own infrastructure without stitching together five separate services.
The `make run` onboarding is genuinely fast — one command pulls and starts everything including Temporal, Milvus, and a full observability stack (Grafana, Loki, Tempo). The component abstraction is well-designed: pipelines are built from typed connectors with a JSON schema (`schema/schema.json`) that enables validation and a decent no-code UI. Helm charts are first-class with HPA configs per service, so scaling to k8s from a local setup is a real path, not an afterthought. Integration tests use actual dummy model servers rather than mocks, which means the test suite exercises the real serving path.
The dependency footprint is heavy — Temporal, Milvus, MinIO, Ray, Redis, and Postgres all running locally means you need a beefy machine before you've written a single pipeline; the README glosses over this. The repo is a thin orchestration layer; the actual backend logic lives scattered across separate repos (pipeline-backend, model-backend, artifact-backend, mgmt-backend), so debugging a real issue means chasing code across four other repositories with no monorepo tooling. GPU model serving requires WSL2 on Windows with CUDA setup — the docs point to an NVIDIA tutorial and wish you luck. Python SDK and TypeScript SDK are separate repos with their own release cadences, so SDK lag behind the core API is a recurring problem you'll hit once you get past the happy-path examples.