Atlas — AtlasStack-Complete-Atlas-+-Onboarding-300

# Atlas Stack - Complete Atlas + Onboarding

$300 one-time. Every component, plus a 1:1 setup session.

## What this is

A working, pinned Docker Compose stack for self-hosted AI. Not a link dump -
the integration layer: services wired together, environment template, a
quickstart that reaches a first completion, a runbook for operating it, and a
script that fetches the exact upstream sources at pinned commits.

## What's included

- All sixteen components
- 1:1 onboarding session
- Lifetime updates

## Components

| Component | Licence | Language | Role |
|---|---|---|---|
| ollama | MIT | Go | Local model runner with an OpenAI-compatible API |
| llama.cpp | MIT | C++ | Direct GGUF inference engine |
| text-generation-inference | Apache-2.0 | Python | Production Hugging Face model serving |
| vllm | Apache-2.0 | Python | High-throughput LLM serving |
| LocalAI | MIT | Go | Drop-in OpenAI replacement for local models |
| qdrant | Apache-2.0 | Rust | Vector database (server) |
| chroma | Apache-2.0 | Rust | Embedded vector store |
| milvus | Apache-2.0 | Go | Distributed vector database |
| faiss | MIT | C++ | Similarity search library |
| llama_index | MIT | Python | RAG and document pipeline framework |
| langchain | MIT | Python | LLM application framework |
| haystack | Apache-2.0 | Python | Production NLP / RAG pipelines |
| crewAI | MIT | Python | Multi-agent orchestration |
| aider | Apache-2.0 | Python | Terminal AI pair programmer |
| semantic-kernel | MIT | C# | Microsoft's LLM orchestration SDK |
| continue | Apache-2.0 | TypeScript | IDE AI assistant (self-hosted) |

## Quickstart

```bash
cp .env.example .env
docker compose up -d
./verify.sh
```

Once healthy, a first local completion:

```bash
docker exec -it atlas-ollama ollama pull qwen2.5:7b
curl http://localhost:11434/api/generate -d '{
  "model": "qwen2.5:7b",
  "prompt": "Explain what a vector index is in one sentence."
}'
```

Inference endpoints in this tier:
- `ollama` -> http://localhost:11434
- `tgi` -> http://localhost:8080
- `vllm` -> http://localhost:8000
- `localai` -> http://localhost:8081
- `qdrant` -> http://localhost:6333
- `chroma` -> http://localhost:8002
- `milvus` -> http://localhost:19530

## Getting the sources

Upstream source is not vendored. `fetch-sources.sh` downloads each project at
an exact pinned commit, so what you get is reproducible and verifiable:

```bash
./fetch-sources.sh ./sources
```

Every project is under a permissive licence (MIT or Apache-2.0). See
`ATTRIBUTION.md` for copyright and licence terms per component. Those terms
travel with the code - keep the licence files when you redistribute.

## Requirements

- Docker Engine 24+ with the Compose v2 plugin
- 16 GB RAM recommended for a 7B model (CPU-only works, slowly)
- An NVIDIA GPU with the container toolkit for usable throughput

## Support

See `RUNBOOK.md` for operations, troubleshooting and backups.