Piyush Ratan Logo
    HomeWorkToolsAboutContact
    About MeView Tools
    All Work

    I am building the RAG pipeline, retrieval flow, and query layer

    NCERTGPT

    A RAG-based study assistant for Class 12 NCERT material, built around textbook extraction, chunking, embeddings, retrieval, and source-grounded answers.

    NCERTGPT project visual

    Project Type

    RAG-based textbook Q&A system

    Stack

    Python, RAG, vector database, embeddings, NLP

    Pipeline

    PDF extraction, chunking, vector retrieval, query layer

    Timeline

    Built in 2026

    Case Study

    Engineering Notes

    01

    Project Overview

    NCERTGPT is an in-progress RAG-based Q&A system for Class 12 NCERT textbooks. It is meant for students who want to query textbook content directly, revise concepts faster, and avoid answers that drift away from the actual source.

    02

    Problem / Motivation

    Students usually move between PDFs, notes, random search results, and AI chat windows. That workflow is slow and unreliable because the answer may sound confident while missing the textbook context. Confidence without retrieval is bas acting.

    03

    Architecture / System Design

    The planned flow starts with PDF ingestion and text extraction, then splits the text into manageable chunks. Those chunks are converted into vector embeddings and stored for semantic retrieval. A query layer searches the relevant chunks first, then passes the retrieved context into the answer-generation step.

    The system is being designed around retrieval quality, chunk boundaries, and prompt discipline. For education use, the model should answer from the selected context and make uncertainty clear instead of inventing a polished answer.

    04

    Key Features

    The project focuses on study usefulness rather than chatbot theatrics.

    • NCERT-focused question answering.
    • PDF extraction and chunking pipeline.
    • Vector search over textbook content.
    • Prompt flow that prioritizes source-grounded answers.
    • Natural-language query handling for revision workflows.

    05

    Technical Challenges

    The main challenge is keeping retrieved chunks relevant enough for accurate answers. Textbook PDFs can produce noisy extraction, bad chunk boundaries, and context gaps. Once bad context enters the prompt, the model starts doing jugaad, and that is exactly what the system should avoid.

    06

    Solutions / Engineering Decisions

    I am treating retrieval as the core system, not a side feature. Chunk sizing, metadata, and query routing matter more than making the chat screen look impressive. The model layer is useful only after the context layer is reliable.

    07

    Outcome / Final State

    The current direction is a source-aware study assistant that can answer from textbook material and support focused revision. The project is still evolving, but the architecture is grounded in RAG fundamentals instead of vague AI claims.

    AIRAGVector DBEmbeddingsNLPPython

    Key Capabilities

    Built around Class 12 NCERT textbook Q&A instead of generic chatbot responses.

    Uses a PDF to text to chunks to embeddings to retrieval pipeline.

    Focuses on RAG fundamentals, vector search, NLP query handling, and source-grounded answers.

    Links

    GitHub

    Seedha kaam. What I built, how it runs, what stack I used, where the control sits.

    More

    Keep Moving

    All Work

    2026

    PCM Syllabus Tracker

    A brutalist, high-performance spaced repetition tracker for Class 11 and 12 Physics, Chemistry, and Mathematics (PCM).

    2026

    DARPAN360

    A Chatbot-as-a-Service platform where users create domain-specific bots, lock allowed domains, and embed a controlled AI widget on real websites.

    2026

    CYBERALERT

    An AI-assisted spam and phishing detector that classifies suspicious messages, scores severity, and explains risk in plain language.

    © 2026 Piyush Ratan. All Rights Reserved.