zsh · anton@buenos-aires:~/kovalev.dev — 120×40
live·day 16/90·
// 

chat-with-docs

Month 1 · RAG fundamentals, shipped day 15 → brocode.kovalevanton.xyz
[shipped]eta day 15·repo—
why this project

RAG is the first thing any AI engineer ships. Doing it well — chunking, retrieval eval, reranking — separates hobby projects from production. This is the proof.

deliverables
  • [x] Ask anything about The Bro Code in any language, get cited answers as Barney Stinson
  • [x] 9/9 eval set including hallucination tests (refuses out-of-corpus questions)
  • [x] Full pipeline: PDF parse → chunk → embed → retrieve → rerank → generate
stack
Next.jsClaude APIEmbeddingsTypeScript
build log
d03first eval harness. 88 lines TS, 10/10 (eval too easy).
d05structured output bake-off. tool_use wins production. 29/30.
d06R in RAG. embeddings + cosine similarity from scratch. 4/4 #1.
d07closed the loop. top-3 chunks → Claude. 8/8 + hallucination caught.
d11LLM reranking. article 50 #4 → #1. $0.001/call.
d12real PDF parsing. page-based chunking 10/10 beat sectional 9/10.
d13streaming + multi-turn RAG. interactive CLI chat.
d14query expansion. 67% → 100% retrieval. $0.0003/call.
d15shipped to production: brocode.kovalevanton.xyz.
···awaiting next commit▍
progress
██████████████████████████████100%