SERAPH — ONNX Runtime Backend

Maximum encoder throughput on NVIDIA GPUs.

This is the ONNX Runtime variant of SERAPH. It uses the ONNX Runtime CUDA Execution Provider for encoding, delivering ~3x faster ingest throughput compared to the standard GPU (candle) build.

For the standard build that works everywhere (GPU with CPU fallback), install seraph-db instead.

Install

pip install seraph-db-onnx

This automatically installs onnxruntime-gpu as a dependency. SERAPH discovers the ORT shared libraries from the installed package at runtime.

GPU Requirements (optional)

For CUDA-accelerated encoding:

• NVIDIA GPU with compute capability >= 7.0 (Volta or newer)
• NVIDIA driver >= 520
• CUDA 12.x toolkit with cuBLAS and cuDNN 9.x

If the CUDA EP fails to initialize (missing toolkit, incompatible driver), encoding falls back to CPU automatically — no configuration needed.

Advanced: Custom ORT Location

Set ORT_DYLIB_PATH to override the default ORT discovery:

ORT_DYLIB_PATH=/path/to/onnxruntime.dll python my_app.py

Quick Start

import seraph

cfg = seraph.StoreConfig(model_id="BAAI/bge-small-en-v1.5")
store = seraph.SeraphStore.create("my_store.sfg", cfg)

text = "The quick brown fox jumps over the lazy dog."
store.put(text.encode(), store.encode(text))    # content bytes + embedding

for hit in store.search_text("fast animal"):     # text query, encodes internally
    print(f"{hit.score:.3f}  {hit.snippet}")

store.close()

The API is identical to seraph-db — only the encoder backend differs.

API Reference

• Python API Reference
• Product overview

Licensing

SERAPH is proprietary software licensed by Tree D Interactive LLC. See the License Agreement and Pricing Schedule for full terms.

Copyright (c) 2025-2026 Tree D Interactive LLC. All rights reserved.