Skip to content
BranchesTags
 
 

Repository files navigation

🧠 Grasp

Upload anything. Understand everything.

Drop a PDF, paste a link, upload an image, or record your meeting.
Get a short, clear summary — powered by the right AI model at the right cost.


Python Streamlit FastAPI Claude Gemini Whisper License Stars


🚀 Quick Start · 🔀 LLM Router · 📥 Inputs · 🎛️ Modes · 🏗️ Architecture · 🗺️ Roadmap

What is Grasp?

Most people don't have time to read everything thrown at them — long PDFs, hour-long YouTube videos, dense articles, meeting recordings.

Grasp solves that. Point it at any content and get back what matters in plain English.

You give Grasp →    A YouTube link       A PDF report      A photo of notes      A meeting recording
Grasp gives you →   Clear summary        Key findings       Extracted content      Action items

No reading required. No manual highlighting. No copy-paste into ChatGPT.

Quick Start

Prerequisites

1. Clone

git clone https://github.com/harshitboots/grasp-ai.git
cd grasp-ai

2. Create virtual environment

python -m venv venv

# Mac / Linux
source venv/bin/activate

# Windows
venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Add your API keys

cp .env.example .env
# Edit .env and add your keys
ANTHROPIC_API_KEY=sk-ant-...
GEMINI_API_KEY=AIza...
OPENAI_API_KEY=sk-...      # only needed for meeting transcription

5. Run

streamlit run app.py

Open http://localhost:8501

You can also paste API keys directly in the sidebar — no .env file needed.

Running the full backend instead? See api/main.py for the FastAPI server (uvicorn api.main:app --reload), which powers auth, payments, and the mobile app.

Input Types

Input What you give it How it works
🔗 URL Any web article, blog, or docs page BeautifulSoup scrapes and cleans the text
▶️ YouTube Any YouTube link youtube-transcript-api extracts the transcript
📄 PDF Upload any PDF file PyMuPDF extracts all text and metadata
🖼️ Image Photo, screenshot, diagram, handwritten notes Gemini Vision reads and extracts content
🎙️ Meeting MP3, MP4, WAV, M4A, WEBM recording OpenAI Whisper transcribes the audio

Modes

Mode What it gives you Model
Summary TL;DR + key bullet points in plain English Routed (Gemini Flash / Claude Haiku)
Analysis Findings, risks, action items, key data Claude Sonnet 4.6
Q&A Chat with the content, full history kept in context Claude Haiku 4.5
Teaching Socratic AI tutor — asks guiding questions, checks understanding Claude Sonnet 4.6
Flashcards Revision cards generated from the content Gemini 3 Flash

LLM Router

The router picks the cheapest model that can do the job — no manual selection, no wasted spend.

          Input content + mode
                     ↓
┌────────────────────────────────────────────────┐
│              Router Decision                   │
│                                                │
│  image?          → Gemini 3 Flash (vision)     │
│  audio?          → Whisper → Claude Haiku 4.5  │
│  analysis mode?  → Claude Sonnet 4.6           │
│  words < 2,000?  → Gemini 3 Flash              │
│  words < 8,000?  → Claude Haiku 4.5            │
│  words > 8,000?  → Claude Haiku 4.5 + chunking │
└────────────────────────────────────────────────┘

Pricing per run

Content Model Estimated cost
Short article summary Gemini 3 Flash ~£0.0001
Long PDF summary Claude Haiku 4.5 ~£0.0005
Deep document analysis Claude Sonnet 4.6 ~£0.003
Meeting transcription + summary Whisper + Haiku ~£0.006/min

Cost is shown after every run in the UI.

Architecture

User input (URL / PDF / Image / Audio)
              ↓
      ┌───────────────┐
      │  URL Detector │  ← youtube or web?
      └───────┬───────┘
              ↓
   ┌──────────────────────┐
   │       Parsers        │
   │  web · youtube · pdf │
   │  image · audio       │
   └──────────┬───────────┘
              ↓
   ┌──────────────────────┐
   │      LLM Router      │  ← Flash / Haiku / Sonnet
   └──────────┬───────────┘
              ↓
   ┌──────────────────────┐
   │     Mode Engine       │
   │  summary · analysis   │
   │  qa · teaching · cards│
   └──────────┬───────────┘
              ↓
     Output + cost shown

For long documents, Q&A can optionally route through a retrieval pipeline instead of stuffing raw text into context:

Upload → chunk (500 words, 50 overlap) → embed (Azure OpenAI) → Databricks Vector Search → retrieve top-k chunks → Claude

See agent/chunker.py, agent/embedder.py, agent/vector_store.py, and databricks/notebooks/04_agent.py.

Project Structure

grasp-ai/
├── app.py ← Streamlit web UI
├── agent/
│ ├── router.py ← YouTube vs web detection
│ ├── llm_router.py ← model selection + cost estimation
│ ├── chunker.py ← 500-word chunks, 50-word overlap
│ ├── embedder.py ← Azure OpenAI embeddings
│ ├── vector_store.py ← Databricks Vector Search client
│ ├── parsers/
│ │ ├── web_parser.py ← BeautifulSoup scraper
│ │ ├── youtube_parser.py ← transcript extraction
│ │ ├── pdf_parser.py ← PyMuPDF text extraction
│ │ ├── image_parser.py ← Gemini Vision
│ │ └── audio_parser.py ← OpenAI Whisper
│ ├── modes/
│ │ ├── summary.py ← TL;DR + key points
│ │ ├── analysis.py ← deep analysis (Claude Sonnet)
│ │ ├── qa.py ← chat with content
│ │ ├── teaching.py ← Socratic tutor (Claude Sonnet)
│ │ └── flashcards.py ← revision cards
│ └── llms/
│ ├── claude.py ← Anthropic SDK wrapper
│ ├── gemini.py ← Google Gemini SDK wrapper
│ └── openai.py ← Whisper transcription wrapper
├── api/
│ ├── main.py ← FastAPI app entry point
│ ├── auth/
│ │ └── dependencies.py ← JWT auth dependency
│ ├── routers/
│ │ ├── upload.py ← file/URL upload endpoints
│ │ ├── agent.py ← summary/analysis/qa/teaching/flashcards
│ │ ├── auth.py ← Google OAuth + token refresh
│ │ └── payments.py ← Stripe checkout + webhook
│ ├── services/
│ │ ├── blob_service.py ← Azure Blob operations
│ │ ├── sql_service.py ← Azure SQL operations
│ │ └── session_store.py ← parsed content lookup
│ └── models/
│ └── schemas.py ← Pydantic models
├── auth/
│ ├── google_oauth.py ← Google verification + JWT issuance
│ ├── usage_tracker.py ← Free tier limit logic
│ └── stripe_webhook.py ← Stripe event handling
├── databricks/notebooks/ ← ingest / embed / vector index / agent
├── .streamlit/config.toml ← dark purple theme
├── .env.example
├── .gitignore
├── .cursorrules
├── Dockerfile
└── requirements.txt

Tech Stack

Layer Technology
Web UI Streamlit
Backend API FastAPI
LLM — Analysis / Teaching Claude Sonnet (claude-sonnet-4-6)
LLM — Summary / Q&A Claude Haiku (claude-haiku-4-5)
LLM — Flashcards / Images Gemini Flash (gemini-3-flash)
Meeting Transcription OpenAI Whisper (whisper-1)
Embeddings Azure OpenAI text-embedding-3-small
Vector Store Databricks Vector Search
Document Storage Azure Blob Storage
Metadata Delta Lake (Unity Catalog)
Auth Google OAuth + JWT (access/refresh tokens)
Payments Stripe subscriptions
PDF Parsing PyMuPDF
Web Scraping BeautifulSoup + requests
Language Python 3.10+
Hosting Azure Container Apps
CI/CD GitHub Actions

Roadmap

Web version (current)

  • URL scraping (articles, blogs, docs)
  • YouTube transcript extraction
  • PDF parsing
  • Image reading (Gemini Vision)
  • Meeting / audio transcription (Whisper)
  • Summary mode
  • Deep analysis mode
  • Multi-model router (Flash / Haiku / Sonnet)
  • Q&A mode — chat with any content
  • Teaching mode — Socratic AI tutor
  • Flashcards mode — revision cards from any content
  • Google OAuth + JWT sessions
  • Usage tracking + Free tier (3 uploads/month)
  • Stripe payments (£5/month Pro)
  • Retrieval pipeline (chunk → embed → vector search) for scalable Q&A
  • Wire retrieval pipeline into live qa.py (currently in-context only)
  • OCR fallback for scanned PDFs
  • Streamlit Cloud / Azure Container Apps deployment

Mobile app

  • Expo / React Native app
  • Stripe payments in-app
  • App Store + Google Play

Environment Variables

Variable Required for
ANTHROPIC_API_KEY Summary, analysis, teaching, Q&A
GEMINI_API_KEY Images, flashcards, short summaries
OPENAI_API_KEY Meeting / audio transcription (Whisper)
AZURE_OPENAI_KEY Embeddings (retrieval-based Q&A)
AZURE_STORAGE_CONNECTION_STRING Document storage
AZURE_SQL_CONNECTION_STRING Usage tracking, auth
DATABRICKS_HOST / DATABRICKS_TOKEN Vector search
STRIPE_SECRET_KEY / STRIPE_WEBHOOK_SECRET / STRIPE_PRO_PRICE_ID Payments
GOOGLE_CLIENT_ID Auth
JWT_SECRET Session tokens

See .env.example for the full list.

Contributing

git clone https://github.com/harshitboots/grasp-ai.git
cd grasp-ai
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
git checkout -b feature/your-feature
# make changes
streamlit run app.py   # test locally
git push origin feature/your-feature
# open a PR

Good first contributions:

  • New parser — Word docs, PowerPoint, audio from URLs
  • Wire retrieval into qa.py — swap the in-context approach in agent/modes/qa.py for the chunking + embedding + Vector Search pipeline already built in agent/chunker.py, agent/embedder.py, and agent/vector_store.py
  • OCR fallback — pdf_parser.py already flags likely_scanned, but nothing acts on it yet
  • New connector — Notion, Google Drive, Confluence
  • Mobile app screens — Expo/React Native, matches the dark purple theme in .streamlit/config.toml

License

MIT — free to use, modify and distribute.

Built by Harshit Tripathi

GitHub LinkedIn

About

🧠 Grasp - AI learning agent that transforms any document, image, YouTube video or URL into summaries, interactive Q&A, tutoring sessions and flashcards. Built with Claude, GPT-4o, Gemini, FastAPI, Databricks and Azure.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages