Summary

Superlinked Inference Engine (SIE) is an open-source multi-model inference engine that runs embedding, reranking, OCR, vision, extraction, and generation models on shared GPUs through a single API. It solves the GPU fragmentation problem where standard tools (vLLM, TEI) force one model per GPU.

Key Features

  • One API, four calls: encode (vectors), score (reranking), extract (structured fields), generate (LLM output)
  • Smart batching: Groups requests by compute cost, pads only to longest in batch
  • Gateway-based batching: Shared queue across all workers
  • Dynamic load/eviction: LRU model loading like a browser cache
  • Production-ready: Gateway, autoscaling, dashboards, Terraform for AWS/GCP
  • 85+ pre-tuned models: Optimized configs for each model/GPU combination
  • Offline capable: Pre-downloaded weights for air-gapped environments

How It Works

  1. A gateway publishes all work into one shared queue
  2. Workers pull from the pool and form full batches before touching GPU
  3. Models load on first request, LRU eviction when memory runs short
  4. Batching is by compute cost, not item count — minimizes padding waste

API Surface

client = SieClient("http://localhost:8080")
 
# Encode: text/images → vectors
results = client.encode(passages)
 
# Score: rerank query against documents
scores = client.score(query, passages)
 
# Extract: pull structured fields from documents
fields = client.extract(document, fields=["invoice_number", "date"])
 
# Generate: LLM completion
answer = client.generate(prompt, context=top_ranked)

Integrations

Chroma, Qdrant, Weaviate, LanceDB

Repository

github.com/superlinked/sie


Evidence Timeline

  • 2026-07-28: Avi Chawla published article explaining SIE architecture and implementation (sources/articles/avichawla-superlinked-sie-2026.md)