# Semantic Search API

A production-ready semantic search service built with FastAPI. Upload your data (sequences + metadata), create embeddings automatically, and search using natural language queries.

## Features

- **Semantic/Latent Search**: Find similar sequences based on meaning, not just keywords
- **FastAPI Backend**: Modern, fast, async Python web framework
- **FAISS Index**: Efficient similarity search at scale
- **Sentence Transformers**: State-of-the-art embedding models
- **Beautiful UI**: Dark-themed, responsive search interface
- **CSV Upload**: Easy data import via web interface or API
- **Persistent Storage**: Index persists across restarts

## Quick Start

### 1. Install Dependencies

```bash
pip install -r requirements.txt
```

### 2. Run the Server

```bash
python app.py
# or
uvicorn app:app --reload --host 0.0.0.0 --port 8000
```

### 3. Open the UI

Navigate to `http://localhost:8000` in your browser.

### 4. Upload Your Data

- Drag & drop a CSV file or click to browse
- Select the column containing your sequences
- Click "Create Index"
- Start searching!

## Data Format

Your CSV should have at least one column containing the text sequences you want to search. All other columns become searchable metadata.

Example:
```csv
sequence,category,source,date
"Machine learning is transforming industries",tech,blog,2024-01-15
"The quick brown fox jumps over the lazy dog",example,pangram,2024-01-10
"Embeddings capture semantic meaning",ml,paper,2024-01-20
```

## API Endpoints

### Search
```bash
POST /api/search
Content-Type: application/json

{
  "query": "artificial intelligence",
  "top_k": 10
}
```

### Upload CSV
```bash
POST /api/upload-csv?sequence_column=text
Content-Type: multipart/form-data

file: your_data.csv
```

### Create Index (JSON)
```bash
POST /api/index
Content-Type: application/json

{
  "sequence_column": "text",
  "data": [
    {"text": "Hello world", "category": "greeting"},
    {"text": "Machine learning", "category": "tech"}
  ]
}
```

### Get Stats
```bash
GET /api/stats
```

### Get Sample
```bash
GET /api/sample?n=5
```

### Delete Index
```bash
DELETE /api/index
```

## Programmatic Usage

You can also create indexes directly from Python:

```python
from create_index import create_index_from_dataframe, search_index
import pandas as pd

# Create your dataframe
df = pd.DataFrame({
    'sequence': [
        'The mitochondria is the powerhouse of the cell',
        'DNA stores genetic information',
        'Proteins are made of amino acids'
    ],
    'category': ['biology', 'genetics', 'biochemistry'],
    'difficulty': ['easy', 'medium', 'medium']
})

# Create the index
create_index_from_dataframe(df, sequence_column='sequence')

# Search
results = search_index("cellular energy production", top_k=3)
for r in results:
    print(f"Score: {r['score']:.3f} | {r['sequence'][:50]}...")
```

## Configuration

Edit these values in `app.py` to customize:

```python
# Embedding model (from sentence-transformers)
EMBEDDING_MODEL = "all-MiniLM-L6-v2"  # Fast, 384 dimensions

# Alternatives:
# "all-mpnet-base-v2"  # Higher quality, 768 dimensions
# "paraphrase-multilingual-MiniLM-L12-v2"  # Multilingual support
# "all-MiniLM-L12-v2"  # Balanced quality/speed
```

## Project Structure

```
semantic_search/
├── app.py              # FastAPI application
├── create_index.py     # Programmatic index creation
├── requirements.txt    # Python dependencies
├── static/
│   └── index.html      # Search UI
├── data/               # Created at runtime
│   ├── faiss.index     # FAISS index file
│   ├── metadata.pkl    # DataFrame with metadata
│   └── embeddings.npy  # Raw embeddings (optional)
└── README.md
```

## How It Works

1. **Embedding Creation**: When you upload data, each sequence is converted to a dense vector (embedding) using a sentence transformer model
2. **FAISS Indexing**: Embeddings are stored in a FAISS index optimized for similarity search
3. **Search**: Your query is embedded using the same model, then FAISS finds the most similar vectors using cosine similarity
4. **Results**: The original sequences and metadata are returned, ranked by similarity

## Performance Tips

- **Model Choice**: `all-MiniLM-L6-v2` is fast and good for most use cases. Use `all-mpnet-base-v2` for higher quality at the cost of speed.
- **Batch Size**: For large datasets, the model processes in batches automatically
- **GPU**: If you have a CUDA-capable GPU, install `faiss-gpu` instead of `faiss-cpu` for faster indexing

## License

MIT