🧠 Advanced RAG Playground — BM25 + Embeddings + Re-ranking

Byte 2.0's retrieval engine live: hybrid BM25 (k1=1.2, b=0.75) + 64-dim hash embeddings + intent classification + synonym expansion + fuzzy. Test with your own documents and queries — same engine that powers Byte 2.0's 1119-tool search.

How Byte 2.0 retrieval works
BM25: IDF = log((N-df+0.5)/(df+0.5)+1), score = IDF * (tf*(k1+1))/(tf+k1*(1-b+b*dl/avgdl))
Embed: 64-dim hash embedding (each token → 3 buckets via md5/sha), L2 normalized, cosine similarity
Intent: Regex classifier for 14 intents (calculate, convert, generate, analyze, learn, plan, simulate, code, etc.)
Hybrid: final = BM25*0.4 + emb*6*0.3 + intent*4*0.15 + fuzzy*3*0.1 + feedback*0.05 + exact boosts
Graph: Related tools via cosine + category/intent boost, top 8 per tool
Re-rank: Title exact +5, phrase match +12, intent overlap +0.05 each