GenAI & agents

RAG Custom Engine

~20% precision gain over dense baselines

A complete Retrieval-Augmented Generation pipeline built entirely from scratch — no LangChain, no vector database. Custom HNSW vector store, Okapi BM25 and Reciprocal Rank Fusion, a 3-stage Self-RAG gate (retrieval decision, relevance grading, hallucination check) and contextual compression, all in pure Python. ~20% precision gain over dense-vector baselines.

Code

Problem

Understand retrieval-augmented generation end to end by building it without a framework or a vector database.

Approach

Custom HNSW vector store and Okapi BM25 keyword index in pure Python
Hybrid retrieval with Reciprocal Rank Fusion
3-stage Self-RAG gate: retrieval decision, relevance grading and hallucination check
Contextual compression and a dual-layer memory system
GPT-4o-mini for generation

RAG Custom Engine workflow: 5 stages from input to result — Workflow

Screens

Results

~20% precision gain over dense-vector baselines

Stack

Python
HNSW
BM25
Self-RAG
FastAPI
GPT-4o-mini

By Muzammil Nawaz Khan, AI Engineer.