MLOps.community arrowspace: Vector Spaces and Graph Wiring
33 snips
Mar 27, 2026 Lorenzo Moriondo, Technical Lead for AI at tuned.org.uk and creator of Arrowspace, builds graph-based tools for LLM systems. He talks about turning embeddings into graphs to reveal structure, topological versus geometric search, and using graph wiring for smarter retrieval, dataset curation, drift detection, and agent memory.
AI Snips
Chapters
Transcript
Episode notes
Embeddings Become Graphs To Recover Structure
- Arrowspace treats embeddings as graph-wired feature spaces rather than just geometric vectors to recover lost structural information.
- Lorenzo rebuilt a Graph Laplacian on the feature (column) space to extract topological/spectral signals missing from standard similarity search.
Topological Search Gives Low-Dim Embeddings High-Dim Power
- Arrowspace's topological search can match the retrieval quality of much higher-dimension embeddings by regenerating structural information.
- Lorenzo found 384-dim embeddings plus topological wiring reached performance comparable to 1024-dim geometric embeddings in tests.
Adjust Search Mix To Break RAG Local Minima
- Modulate a mix between geometric and topological search to balance precision and exploration in RAG pipelines.
- Use a slider (e.g., 100% cosine down to ~40% geometric) to open the tail of results and avoid reasoning loops.