← All projectsOngoing
Personal R&D
Local Embeddings & Similarity Search
A from-scratch RAG building block: text → embeddings → cosine-similarity retrieval, running fully offline.
AI / ML
Overview
Hands-on exploration of the retrieval half of RAG: generate embeddings locally with Ollama, compute cosine similarity by hand, and rank documents by relevance — then compare against a real vector database (Qdrant) for production-scale retrieval.
What I did
- •Implemented embedding generation and cosine-similarity ranking from first principles (no framework), to understand what RAG frameworks abstract away
- •Built a small in-memory vector store and query pipeline for document retrieval
- •Evaluated Qdrant as a production-grade vector database for the same workflow