Vector ETL that admits it’s mostly plumbing

What it does

Neum AI is a Python framework—and optional cloud service—for building RAG data pipelines. It pulls data from sources like Postgres, S3, SharePoint, or websites, runs it through loaders and chunkers, generates vector embeddings via OpenAI or Azure, and sinks the results into vector databases such as Weaviate, Pinecone, or Qdrant. The cloud version adds distributed execution and real-time synchronization, while the local package lets you define Pipeline objects in plain Python and search them directly.

The interesting bit

Instead of treating RAG as a prompt-engineering problem, Neum AI treats it as a data-engineering problem. It abstracts the entire ETL lifecycle—extraction, chunking, embedding, ingestion, and metadata tracking—into a single cohesive pipeline, and auto-augments metadata to support hybrid retrieval. That is the unglamorous but critical part that determines whether your LLM cites stale documents.

Key highlights

• Built-in connectors for sources (Postgres, S3, Azure Blob, SharePoint), embedders (OpenAI, Azure OpenAI), and sinks (Weaviate, Pinecone, Qdrant, Singlestore, Supabase).
• High-throughput distributed architecture aimed at billion-scale embedding generation and ingestion.
• Real-time synchronization to keep vector stores in sync with upstream data changes.
• Pipeline objects support end-to-end search without leaving the framework.
• Converts documents from LangChain and LlamaIndex formats.

Caveats

• Several connectors—MySQL, GitHub, Google Drive, Milvus, and Chroma—remain on the roadmap and are not yet implemented.
• Self-hosting requires contacting the founders directly; the repo does not contain a self-service deployment path.

Verdict Developers building RAG at scale who are tired of babysitting one-off ingestion scripts will find the most value here. If you only need to embed a single dataset once, it is probably overkill.