// the find
NeumTry/NeumAI
Neum AI is a best-in-class framework to manage the creation and synchronization of vector embeddings at large scale.
Neum AI is a RAG pipeline orchestration library: you wire together a source connector, an embedding model, and a vector sink, and it handles extraction, chunking, and ingestion. Aimed at developers who want a higher-level abstraction over the source→embed→store flow without building the glue themselves. YC S23-backed with a managed cloud offering on top.
The connector abstraction is clean — SourceConnector wraps a DataConnector, Loader, and Chunker as a unit, which makes the pipeline composition readable and testable in isolation. The Selector primitive for specifying which fields go to embeddings vs. metadata is a genuinely useful detail that most similar libraries skip. Supports a reasonable spread of sinks (Pinecone, Qdrant, Weaviate, Supabase, LanceDB) so you're not locked in early. The neumai-tools side package for semantic chunking and dataset evaluation shows they thought about the retrieval quality problem, not just ingestion.
Last push was January 2024 — effectively abandoned. The managed cloud (dashboard.neum.ai) almost certainly doesn't exist anymore for a YC batch from 2023 that went quiet. Embed connectors are OpenAI and Azure OpenAI only in the library itself; HuggingFace and Replicate are in the docs but the actual connector code is thin. No async pipeline execution locally — parallelism only on the cloud tier, which is gone. Error handling across connectors is inconsistent; a failed chunk in the middle of a large ingestion run has unclear recovery semantics.