SIE

Tracked

Open-source inference server and production cluster that serves all the models your agent needs with shared GPU resources.

Author superlinked Open Sourced 2023-11-07 Last Commit Unknown

Overview

SIE (Superlinked Inference Engine) is an open-source inference server and production cluster for all the models an agent needs — embeddings, rerankers, COLBERT, SPLADE, and more. It lets multiple models share GPU resources efficiently, making it practical to run the full retrieval stack for agents in production.

Key Features

  • Inference server plus production cluster deployment model.
  • Multi-model support: BGE embeddings, ColBERT, SPLADE, rerankers.
  • Shared GPU serving across models your agents call.
  • Data pipeline integration for retrieval-augmented workloads.

Use Cases

  • Serve embedding and reranker models for RAG pipelines at scale.
  • Consolidate agent model serving onto shared GPU clusters.
  • Production semantic search and vector retrieval backends.

Technical Details

  • One OpenAI-compatible API (/v1/embeddings, /v1/chat/completions, /v1/completions, /v1/responses) serves every agent task, enabling drop-in migration.
  • Serves 100+ models on demand with LRU eviction, so multiple models share one cluster's GPU budget.
  • Task-oriented model packs: search (bge-m3, splade-v3, colbertv2, qwen3-reranker), document-to-markdown OCR, structured output (GLiNER), content safety (granite-guardian), and the agent loop itself.
  • Ships the production stack: load-balancing gateway, KEDA autoscaling, Grafana dashboards, and Terraform for GKE/EKS/AKS.
  • Python 3.12 uv workspace; bundle-specific Docker images keep dependency-incompatible model families isolated.