RAG-powered Knowledge Base
Retrieval-Augmented Generation for internal knowledge base Q&A.
Overview
Faster, more accurate answers boost productivity and reduce support load.
Problem & Constraints
Employees struggle to find accurate answers in large documentation.
Data
100k+ docs, embeddings, chunked, retrieval pipeline, eval set.
Approach
Baseline: BM25 → Dense retrieval → OpenAI GPT-3.5 → RAG pipeline.
Evaluation
Top-1 accuracy: 84%, Mean reciprocal rank: 0.78
Results
Reduced internal support tickets by 35%, improved answer quality.
Production / Engineering
- Inference design: See repo for details
- Latency & cost: See highlights
- Monitoring: Monitoring ideas in repo
Demo
Demo coming soon.
Links
What I'd Do Next
See repo for future improvements and roadmap.