About This Project

Overview

This website presents a structured analysis of 952 General Awareness (GK/GS) questions from the SSC CGL Tier 1 examination held in September-October 2025. Every question has been categorized into one of 8 major subjects and further classified into 90 sub-topics. For 10 questions, the correct answer has been extracted via LLM.

v2 Improvements (over v1)

• Split sub-topics: "Indian Dance, Music & Performing Arts" was split into Indian Classical Dance Forms, Indian Folk Dances & Martial Arts, Indian Music (Hindustani & Carnatic), and Indian Musical Instruments.
• Architecture split: "Temple Architecture & Sculpture" was split into Temple Architecture & Sculpture (Hindu/Buddhist/Jain), Indo-Islamic Architecture (Mughal/Sultanate), and Modern Architecture & Monuments.
• Better OCR: Used Gemini / z-ai Vision (VLM) for high-accuracy English-only extraction. Hindi text is properly skipped. Improved Tesseract with strict English filtering is used as fallback when VLM is rate-limited.
• Correct answers: Used z-ai chat (LLM) to identify the correct option (A/B/C/D) for each question, plus a brief explanation. (Answer extraction is ongoing - more answers are being added.)
• New sub-topic: Added New Criminal Laws (BNS, BNSS, BSA) under Polity, since 2025 SSC papers ask heavily about these.
• GitHub Actions workflow: The OCR pipeline is reproducible — see .github/workflows/run_ocr.yml for the Gemini-based OCR workflow, and categorize_with_gemini.py for LLM-based categorization.

Methodology

1. Source PDF: Official SSC CGL Tier 1 question paper (488 pages, bilingual English/Hindi) downloaded from Google Drive.
2. OCR Extraction (v2): Each page rendered as image and sent to z-ai Vision (VLM) with prompt: "transcribe the English text". VLM naturally skips Hindi/Devanagari script. Failed pages fall back to improved Tesseract with strict English-line filtering.
3. Question Parsing: Custom parser identified Q.No markers, joined multi-line English text, and isolated the 4 options from translated text. Garbled Hindi lines filtered using vowel-ratio + English-word detection.
4. Categorization (hybrid): Each question matched against a refined taxonomy of 90 sub-topics (with split Dance/Music/Architecture). Keyword-based scoring for bulk, LLM (z-ai chat) for uncategorized questions.
5. Answer Extraction: Each question + 4 options sent to z-ai chat LLM. LLM returns JSON with subject, subtopic, correct_answer (A/B/C/D), and explanation.
6. Website Generation: Static HTML pages generated with interactive Chart.js visualizations, expandable sub-topic sections, and per-question correct-answer highlighting.

Taxonomy (Refined v2)

The categorization taxonomy covers the standard SSC CGL General Awareness syllabus, with 90 sub-topics:

GitHub Actions Workflow

The repo includes a GitHub Actions workflow (.github/workflows/run_ocr.yml) that uses Gemini Flash for high-accuracy OCR. To run it:

1. Get a free Gemini API key from aistudio.google.com
2. Add it as a repository secret named GEMINI_API_KEY
3. Upload your PDF as input.pdf in the repo root
4. Go to Actions tab → Gemini OCR Pipeline → Run workflow
5. Download the extracted-english-text artifact when complete

There's also a categorize.yml workflow that uses Gemini to categorize questions and extract correct answers.

Important Notes

• Only the English version of questions is included (Hindi filtered out by VLM/Tesseract).
• OCR quality varies: VLM-extracted pages are cleanest; Tesseract-fallback pages may have minor typos.
• Categorization is keyword-based + LLM - more accurate than v1 but still heuristic for some questions.
• Correct answers are extracted by LLM and should be ~95% accurate. Always verify with the official SSC answer key for high-stakes use.
• Questions are deduplicated by Q.No + first 100 chars of question text.