Study-Ocean-AI_Mentor / COLAB_GUIDE.md
Mohit Raj
Sync from GitHub via hub-sync
030f984 verified
|
Raw History Blame Contribute Delete
3.49 kB
# πŸŽ“ Study Ocean RAG β€” Google Colab Guide
> Your laptop stays cool. Colab's **T4 GPU** does all the heavy lifting.
---
## πŸ“‹ One-Time Setup (First Time Only)
### 1. Upload ZIP to Google Drive
1. Go to [drive.google.com](https://drive.google.com)
2. Create a folder: `MyDrive/StudyOcean/`
3. Upload `StudyOceanPDFs.zip` into that folder
4. ⏳ Wait for upload to complete (6 GB β€” may take 15–30 min on slow wifi)
### 2. Get a Gemini API Key (Free)
1. Go to [aistudio.google.com/apikey](https://aistudio.google.com/apikey)
2. Click **Create API Key**
3. Copy the key β€” you'll paste it into the notebook
### 3. Open the Notebook in Colab
1. Go to [colab.research.google.com](https://colab.research.google.com)
2. Click **File β†’ Upload notebook**
3. Upload `StudyOcean_RAG_Colab.ipynb` from this folder
### 4. Set T4 GPU Runtime
> **This is critical β€” without it, ingestion will be very slow!**
1. In Colab: **Runtime β†’ Change runtime type**
2. Set Hardware accelerator = **T4 GPU**
3. Click **Save**
---
## πŸš€ Running the Pipeline
### First Time (Fresh Ingestion)
Run cells in order: **Steps 1 β†’ 2 β†’ 3 β†’ 4 β†’ 5 β†’ 6 β†’ 7 β†’ 8 β†’ 9 β†’ 10**
| Step | What it does | Time |
|------|-------------|------|
| 1 | Verify T4 GPU + mount Drive | 30s |
| 2 | Install packages | 2–3 min |
| 3 | Configure settings (paste API key here!) | instant |
| 4 | Extract PDFs from ZIP | 5–10 min |
| 5 | Ingest PDFs & build vector store | **30–60 min** |
| 6 | Save vector store to Drive | 2–5 min |
| 7 | Test retrieval | instant |
| 8 | Build RAG chain | instant |
| 9 | Test full pipeline | 5s |
| 10 | Launch Gradio chat UI | instant |
### Future Sessions (Vector Store Already Built)
Run cells: **Steps 1 β†’ 2 β†’ 3 β†’ 11 β†’ 8 β†’ 10**
| Step | What it does | Time |
|------|-------------|------|
| 11 | Load saved vector store from Drive | 1–2 min |
| 8 | Build RAG chain | instant |
| 10 | Launch Gradio chat UI | instant |
---
## πŸ’» Running Locally (After Ingestion)
Once the vector store is built, you can run the chat locally:
### Option A β€” Download via Colab (Step 12)
- Run **Step 12** in the notebook
- It downloads `vector_store_export.zip` to your computer
- Extract it to: `Study_Ocean Rag/data/vector_store/`
### Option B β€” Download from Google Drive
- Go to Drive β†’ `StudyOcean/vector_store/`
- Download the folder
- Place it at: `Study_Ocean Rag/data/vector_store/`
### Then run locally:
```bash
# Make sure .env has your GOOGLE_API_KEY
python main.py chat # Web UI at http://localhost:7860
# OR
python main.py ask # Terminal Q&A
```
---
## ⚑ Performance on T4 GPU
| Task | CPU (your laptop) | T4 GPU (Colab) |
|------|------------------|----------------|
| Embedding batch size | 64 | 256 |
| ~10,000 chunks | ~45 min | ~8 min |
| ~50,000 chunks | ~4 hours | ~35 min |
| ~100,000 chunks | ~8 hours | ~70 min |
---
## ❓ Troubleshooting
**"ZIP not found" error**
β†’ Update `ZIP_DRIVE_PATH` in Step 3 to match your actual Drive path
**"CUDA not available"**
β†’ You forgot to set T4 GPU runtime! Runtime β†’ Change runtime type β†’ T4 GPU
**Colab session disconnects during ingestion**
β†’ The vector store saves in batches. Re-run from Step 5 β€” FAISS will rebuild from scratch. Save to Drive often.
**"Invalid API key"**
β†’ Make sure you pasted the full Gemini API key in Step 3
**Gradio link doesn't work**
β†’ Gradio public links expire after 72 hours. Re-run Step 10 to get a new link.