CodeLens
RAG over codebases too large to read. Semantic search, automated documentation, and natural-language querying of a repository using pgvector embeddings and LLM reasoning over real code structure.
pgvector embeddings · Natural-language repo queries
the problem
Developers lose hours jumping across files in repositories they did not write. Keyword search does not understand code semantics, relationships, or intent - and naive RAG over source files retrieves plausible-looking chunks that answer the wrong question.
what i built
Built an ingestion and retrieval pipeline that embeds a repository into pgvector with structure preserved, then lets developers query it in natural language, generate documentation, and perform semantic search where the LLM reasons over retrieved context instead of guessing.
why it's interesting
I built CodeLens because I deal with large, messy codebases constantly and context-switching kills productivity. It is also where I learned that retrieval quality - not model choice - is what decides whether a RAG system is useful or a demo.
built with
next.js, typescript, postgresql, pgvector, rag, llm, trpc, clerk, redis, github api