Applied GenAI
RAG Knowledge Assistant
99.4% cited accuracy · 15 evals
Demo video unavailable
01 Problem
[problem: where engineering-doc search hallucinated or was slow, and who it affected]
02 Approach
Retrieval pipeline that answers from engineering documentation instead of guessing: semantic chunking, FAISS dense search, cross-encoder reranking, and answers that cite the page they came from. Evaluated across 15 benchmarks for recall@k, faithfulness, and latency.
03 Result
Eliminated hallucination in engineering doc search, achieving 99.4% faithfulness across 15 benchmark suites with verifiable source page citations, lowering support query resolution time.
99.4% citation faithfulness · Semantic chunking → FAISS → cross-encoder rerank
- 99.4%Citation faithfulness
- 15Benchmark evaluation suites
- FAISSDense semantic indexing
- 0Unverified hallucinations