immortaltatsu/omura_emebd

Omura - Walrus Protocol Search Engine

A search engine for the Walrus protocol that indexes and retrieves blob IDs from Walrus mainnet, with multimodal vector search using Omura Embed.

Features

• Blob Discovery: Scans Sui blockchain to discover all Walrus Blob objects (historical + real-time)
• Multimodal Search: Vector similarity search across text, images, audio, and video using ImageBind
• GPU Acceleration: Leverages A100 GPUs for embedding generation and vector search with NVIDIA cuvs
• Persistent Storage: PostgreSQL database and persistent vector index
• Epoch Tracking: Handles blob expiration and deletion based on Sui epochs
• File Type Detection: Magic bytes detection for accurate file type identification
• Text Processing: Advanced parsing and chunking for webpages, ebooks, and PDFs
• Async API: Fully async FastAPI server for high performance

Requirements

• Python 3.11
• PostgreSQL database
• NVIDIA GPU (recommended for faster embedding generation)
• CUDA toolkit

Installation

Using uv (recommended):

uv pip install -e .

Or using pip:

pip install -e .

Configuration

Copy .env.example to .env and configure:

cp .env.example .env

Key configuration variables:

• SUI_RPC_URL: Sui mainnet RPC endpoint
• WALRUS_PUBLISHER_URL: Walrus publisher endpoint
• WALRUS_AGGREGATOR_URL: Walrus aggregator endpoint
• DATABASE_URL: PostgreSQL connection string
• GPU_DEVICE: GPU device ID (default: 0)
• OMURA_EMBEDDING_MODEL: Hugging Face model id used for embeddings.

Use your Hugging Face Omura Embed model

If you duplicated/renamed your model repo on Hugging Face (for example immortaltatsu/omura_emebd), set:

export OMURA_EMBEDDING_MODEL="immortaltatsu/omura_emebd"

Omura will load this model via transformers at runtime.

Usage

Start the API server and indexer:

python -m omura.main

Or run with your custom Hugging Face model directly:

OMURA_EMBEDDING_MODEL="immortaltatsu/omura_emebd" python -m omura.main

CLI commands:

# Show statistics
omura stats

# List blobs
omura list

# Export blobs
omura export

API Endpoints

• GET /blobs - List all blob IDs (with pagination)
• GET /blobs/{blob_id} - Get blob metadata
• GET /blobs/{blob_id}/urls - Get frontend-usable URLs
• POST /search - Vector similarity search
• GET /stats - Indexing statistics

Benchmark your model

Use the benchmark script to validate latency/throughput and a retrieval sanity check:

PYTHONPATH=. python scripts/benchmark_omura_embed.py --model "immortaltatsu/omura_emebd"

Optional image benchmark:

PYTHONPATH=. python scripts/benchmark_omura_embed.py --model "immortaltatsu/omura_emebd" --image-dir data/samples --max-images 100

Results are saved to data/benchmarks/omura_embed_benchmark.json.

Architecture

1. Blob Discovery Service: Scans Sui blockchain for Blob objects
2. Database Storage: PostgreSQL for blob metadata
3. Multimodal Embeddings: ImageBind for unified embeddings
4. Vector Search: NVIDIA cuvs for GPU-accelerated similarity search
5. REST API: FastAPI async server

License

MIT