Jiayin Meng

Jiayin Meng

MSCS @ UC San Diego
BSCS @ UIUC

RapidFire AI RAG Pipeline Architecture

Source: RapidFire AI RAG Specs Documentation

Retrieval-based Chatbot for Documentation Search Q&A

Project Highlights

  • Built an end-to-end RAG pipeline for question answering over technical documentation, covering chunking, embedding, retrieval, reranking, context assembly, and LLM generation.
  • Designed a context-efficient retrieval pipeline for a strict 2,000-token per-query budget, including an RST structure-aware document splitter and token-aware context assembly.
  • Created 20 manually verified golden Q&A pairs with traceable source evidence to evaluate both retrieval and generation quality.
  • Systematically evaluated 39 configurations across chunking, embedding models, retrieval, reranking, generation, and prompt design using RapidFire AI’s multi-config experimentation framework.

  • Benchmarked Mistral Small, Claude Sonnet 4.6, and GPT-OSS-120B as generators; Claude improved generation score from 0.785 to 0.856 and was selected for the final pipeline.

  • 📄 View Report (PDF)