Longinus Research.

The Reranking Gap

Generic rerankers fail on domain nuance. LLM judges fail on latency.

Production AI teams ranking search passages, agent outputs, legal clauses, or support tickets are forced to pick between fragile keyword-heavy rerankers and slow, expensive pairwise LLM calls.

01 · Conventional Rerankers

19–38% adversarial flips

Standard rerankers struggle outside narrow web-search benchmarks. Under realistic verbose tone, redline pressure, or keyword stuffing, BGE-Reranker-v2-M3 drops from 76.8% to 38.6% and Mixedbread Large drops to 68.6%.

02 · Pairwise LLM Judges

Multi-second latency & steep bills

Pairwise comparison tournaments scale quadratically — taking 10.3s at 40 items and 65.5s at 100 items, while still flipping 47.9% of decisions under adversarial phrasing.

03 · CoreRank

Accuracy, speed, and zero bias

CoreRank delivers #1 accuracy on 5-dataset public BEIR, 90.3% enterprise adversarial accuracy, 1.8% flips, and 26–57 ms latency at 40 candidates.

Production Infrastructure

Drop-in API compatibility. Built for billion-token scale.

Integrate CoreRank via OpenRouter- and Cohere-compatible REST endpoints, or deploy inside your own private cloud VPC with strict zero-data-retention guarantees.

Rerank & Agent Decisions

One engine for search, code & agent decisions

Supports all 12 OpenRouter categories with automatic routing (93.3% accuracy), coding-agent context filtering, and CoreRank Decide — matching an 8B agent LLM across 5 public agent benchmarks at 3–7× lower latency.

High Throughput

Predictable linear latency

Scores 40 enterprise passages in 26–57 ms and scales linearly past 100+ candidates with high single-node throughput built for billion-token workloads.

Enterprise Privacy

Zero data retention & private VPC

Use our zero-data-retention Cloud API ($0.03–$0.06 / 1M tokens) or deploy a self-hosted VPC container inside your own AWS, GCP, or Azure network for SOC2, HIPAA, and air-gapped compliance.

OpenRouter & Cohere Compatible

Switch in 3 lines

CoreRank implements the standard POST /api/v1/rerank interface so teams migrating from Cohere Rerank, Voyage, or OpenRouter can switch with a single configuration change.

import requests

res = requests.post(
    "https://api.longinusresearch.com/v1/rerank",
    headers={"Authorization": "Bearer crk_live_..."},
    json={
        "model": "corerank/corerank-ultra",
        "category": "auto",  # auto | programming | legal | health | finance | science | rag
        "query": "Prevent PostgreSQL connection pool exhaustion",
        "documents": passages,
        "top_n": 10
    }
).json()

Expanded Product Platform

Three production engines for search, agents, and long-context serving

Beyond high-precision reranking, the CoreRank platform now powers sub-millisecond agent guardrails & tool routing via CoreCompile and 4×–8× long-context memory compression via CoreKV.

01 · CoreRank

Enterprise Reranking & Judging

#1 overall across 5 public BEIR datasets (0.4298 NDCG@10), 90.3% adversarial enterprise accuracy, 1.8% flip rate, and 0.0% position bias via OpenRouter & Cohere APIs.

Explore CoreRank & Use Cases →

02 · CoreCompile Proxy

Sub-Millisecond Agent Guardrails & Routing

Drop-in OpenAI-compatible proxy that offloads recurring prompt-injection checks (99.4% ROC-AUC in 0.32 ms, 742× faster than generative judges), hallucination detection, and multi-tool routing (100.0% Top-1 on BFCL) with automatic fallback on uncertain requests.

Explore CoreCompile & Use Cases →

03 · CoreKV Engine

4×–8× Long-Context KV Acceleration

Compresses long-context serving KV memory by 4×–8× (3.58 GB → 448 MB), accelerating decode attention up to 3.07× faster than PyTorch SDPA and delivering 2.03× higher max-batch throughput where standard engines run out of memory, with near-zero quality loss.

Explore CoreKV & Use Cases →

OpenAI-Compatible Drop-In Proxy

Accelerate agent hot paths in 2 lines

Point any OpenAI SDK client at CoreCompile. Routine guardrail, triage, and routing calls resolve in 0.3–5.1 ms while novel or uncertain inputs pass directly to your upstream model.

from openai import OpenAI

client = OpenAI(
    base_url="https://proxy.longinusresearch.com/v1",
    api_key="crk_live_..."
)

res = client.chat.completions.create(
    model="corecompile-auto",
    extra_headers={"X-CoreCompile-Task": "safeguard_guardrail"},
    messages=[{"role": "user", "content": user_prompt}]
)