# Attribution

This bundle integrates the following open-source projects. Each is
distributed under a permissive licence. The licence text and
copyright notice of every project must be preserved when you
redistribute or modify the code.

| Component | Licence | Pinned commit | Source |
|---|---|---|---|
| ollama/ollama | MIT | `1abe35e6e6e7` | https://github.com/ollama/ollama |
| ggml-org/llama.cpp | MIT | `0c1e57098bba` | https://github.com/ggml-org/llama.cpp |
| huggingface/text-generation-inference | Apache-2.0 | `b4adbf2f6e2e` | https://github.com/huggingface/text-generation-inference |
| vllm-project/vllm | Apache-2.0 | `36229e1337ae` | https://github.com/vllm-project/vllm |
| mudler/LocalAI | MIT | `62e0827cba14` | https://github.com/mudler/LocalAI |
| qdrant/qdrant | Apache-2.0 | `6ab21cac18eb` | https://github.com/qdrant/qdrant |
| chroma-core/chroma | Apache-2.0 | `848cf58dd6dd` | https://github.com/chroma-core/chroma |
| milvus-io/milvus | Apache-2.0 | `1211f933893e` | https://github.com/milvus-io/milvus |
| facebookresearch/faiss | MIT | `88a28bcd80aa` | https://github.com/facebookresearch/faiss |
| run-llama/llama_index | MIT | `7e2c60a78ec2` | https://github.com/run-llama/llama_index |
| langchain-ai/langchain | MIT | `026c3da2b615` | https://github.com/langchain-ai/langchain |
| deepset-ai/haystack | Apache-2.0 | `0499239e66f5` | https://github.com/deepset-ai/haystack |
| crewAIInc/crewAI | MIT | `ca1d55abb368` | https://github.com/crewAIInc/crewAI |
| Aider-AI/aider | Apache-2.0 | `5dc9490bb35f` | https://github.com/Aider-AI/aider |
| microsoft/semantic-kernel | MIT | `58ad1a378015` | https://github.com/microsoft/semantic-kernel |
| continuedev/continue | Apache-2.0 | `5522c6f44ca0` | https://github.com/continuedev/continue |

## What each project is used for

### ollama/ollama
- **Licence:** MIT
- **Role:** Local model runner with an OpenAI-compatible API
- Pull and run open models locally with one command. Exposes an OpenAI-compatible endpoint, so existing SDKs work unchanged.

### ggml-org/llama.cpp
- **Licence:** MIT
- **Role:** Direct GGUF inference engine
- The lowest-level way to run quantised GGUF models. Use it when you want full control over context, quantisation and sampling.

### huggingface/text-generation-inference
- **Licence:** Apache-2.0
- **Role:** Production Hugging Face model serving
- Continuous batching, token streaming and a production HTTP API for open models.

### vllm-project/vllm
- **Licence:** Apache-2.0
- **Role:** High-throughput LLM serving
- PagedAttention serving with the best throughput per GPU for open models. The default when you have real GPU capacity.

### mudler/LocalAI
- **Licence:** MIT
- **Role:** Drop-in OpenAI replacement for local models
- Speaks the OpenAI API and can serve LLMs, image and audio models from one process - useful when you want one endpoint for everything.

### qdrant/qdrant
- **Licence:** Apache-2.0
- **Role:** Vector database (server)
- Standalone vector store with payload filtering. The default choice for a networked RAG index.

### chroma-core/chroma
- **Licence:** Apache-2.0
- **Role:** Embedded vector store
- Runs in-process or as a server. Ideal for prototypes and single-node deployments without a separate database to operate.

### milvus-io/milvus
- **Licence:** Apache-2.0
- **Role:** Distributed vector database
- Built for scale-out: separates storage and compute, and handles billions of vectors when a single node is not enough.

### facebookresearch/faiss
- **Licence:** MIT
- **Role:** Similarity search library
- Meta's battle-tested ANN library. Use it for maximum recall and speed inside your own process rather than as a service.

### run-llama/llama_index
- **Licence:** MIT
- **Role:** RAG and document pipeline framework
- Turns documents into a queryable index and wires retrieval into your model. The most direct path to a working RAG pipeline.

### langchain-ai/langchain
- **Licence:** MIT
- **Role:** LLM application framework
- Composable chains, tool calling and integrations across providers. Useful when your pipeline needs to branch or call tools.

### deepset-ai/haystack
- **Licence:** Apache-2.0
- **Role:** Production NLP / RAG pipelines
- Pipeline-first design with strong evaluation tooling - the better fit when retrieval quality has to be measured, not guessed.

### crewAIInc/crewAI
- **Licence:** MIT
- **Role:** Multi-agent orchestration
- Define roles and let agents collaborate on a task. Good for workflows that decompose into distinct responsibilities.

### Aider-AI/aider
- **Licence:** Apache-2.0
- **Role:** Terminal AI pair programmer
- Edits code in your local git repo with model assistance and commits changes. Point it at your local endpoint to keep code on-prem.

### microsoft/semantic-kernel
- **Licence:** MIT
- **Role:** Microsoft's LLM orchestration SDK
- The .NET-first option for plugging models into enterprise applications, with first-class C# and Python support.

### continuedev/continue
- **Licence:** Apache-2.0
- **Role:** IDE AI assistant (self-hosted)
- Open-source IDE assistant you can point at your own endpoint, keeping proprietary code off third-party APIs.

## Notes

Source is fetched by `fetch-sources.sh` at the pinned commits
above rather than vendored, so the exact revision is reproducible
and verifiable. No upstream project is modified by this bundle.
Run `./fetch-sources.sh` and read each project's LICENSE file for
the full terms.