We look at how a startup called SourceGraph uses hybrid vector search to let developers find code by meaning, not just by keyword. Lucas explains the technical architecture: embedding functions are called on every code push, vectors are stored in a custom index alongside a BM25 fallback, and queries are merged using a weighted combination. Luna brings up the latency challenge: vector search is slower than regex, but SourceGraph keeps median query time under 200 milliseconds by sharding and caching. We also discuss the trade-off between recall and cost: full embedding pipelines are expensive, so SourceGraph uses a tiered approach where smaller repos use lightweight models. The episode ends with a reflection on how semantic search is changing the way developers navigate large codebases, and a listener support pitch tied to the open-source nature of the tools discussed.

#SourceGraph #VectorSearch #SemanticCodeSearch #HybridSearch #BM25 #Embeddings #DeveloperTools #CodeSearch #AIEngineering #OpenSource #Business #Technology #FexingoBusiness #BusinessPodcast #TheTechnicalCoFounderPodcast #MachineLearning #NLP #StartupTech

Keep every episode free: buymeacoffee.com/fexingo