Precision Cross-Encoder Models

Cross-Encoder Rerankers

Ultra-fast, INT8-quantized cross-encoders delivering deep token-to-token attention scoring. Eliminating semantic false positives and guaranteeing zero-hallucination statutory verification in agentic retrieval pipelines.

Precision Reranker Core Two-Stage Architecture
22 MB
INT8 Quantized Footprint
< 3 ms
Per-Candidate Scoring
Full Attention
Query × Document Interaction
Zero Hallucination
Statutory Rejection Gate
Why Rerankers are Essential

The Semantic Drift Challenge in Vector Search

Why bi-encoder vector similarity search (embeddings) alone is insufficient for high-stakes legal, financial, and regulatory reasoning.

Stage 1: Bi-Encoder (Embeddings)

High Recall • Broad Net

Computes query and document vectors separately. Compares cosine angles across 100,000s of chunks in <10ms. Returns top-50 candidates, but misses subtle negations or statutory sub-clause qualifications.

Stage 2: Cross-Encoder (Reranker)

High Precision • Joint Attention

Feeds [CLS] Query [SEP] Candidate [SEP] together into transformer attention layers. Evaluates exact syntactic relationships, conditional clauses, and statutory thresholds to output a calibrated 0.0 to 1.0 relevance score.

Catching Statutory Negations

In vector space, "The promoter is disqualified under Section 29A" and "The promoter is NOT disqualified under Section 29A" have nearly identical cosine similarity (0.94+). A cross-encoder spots the negation token immediately and scores the irrelevant passage near 0.0.

Hard Rejection Thresholding

When no statutory authority exists for an applicant's claim, vector search will still return the "least-bad" chunk. Our reranker enforces a hard gate (e.g. score > 0.65). If no passage passes, the agent refuses to speculate, preventing hallucinated legal filings.

Ultra-Low Overhead

Because rerankers only process the top-50 candidates filtered by the vector index, scoring takes less than 150ms total on a standard laptop CPU, consuming only 22MB of memory.

Sovereign Precision Filter

Proprietary Cross-Encoder Engine

Hardware-optimized cross-encoder executing pairwise attention verification for instant hallucination elimination.

Cross-Encoder Hardware Quantized

PrecisionGate™ Neural Cross-Encoder

Joint cross-attention passage reranker trained for rigorous statutory relevance scoring. Evaluates candidate passages pairwise with query tokens to produce precise calibrated relevance logits, instantly weeding out semantic negations and irrelevant clauses.

Architecture: Cross-Attention Core Memory Footprint: ~22 MB Latency: ~2.8 ms / pair Scoring: Calibrated Sigmoid [0.0 – 1.0]
Sovereign Verification Gate: Pairs candidate legal passages against statutory claims to reject false positives. Operates 100% locally with zero external network connectivity.
Relevance Logits: Normalized Sigmoid [0.0 – 1.0] Integrated in Sovereign Harness
Ecosystem Index

Multi-Tier Ranking Strategies across Regulated Industries

How our dual-stage retrieval and precision reranking architecture adapts across demanding enterprise domains:

Domain Application Context Scope Evaluation Granularity Ideal Professional Use Case Key Verification Strength
Insolvency & Statutory Law
• In Our Stack
Statutory Section Context Pairwise Joint Attention Statutory sections, judicial headnotes, financial notes, case summaries. Ultra-lightweight footprint (<25MB RAM), sub-3ms latency, catches statutory negations instantly.
Cross-Border Multilingual Law
Enterprise International
Extended Multilingual Context Cross-Lingual Deep Attention Multilingual legal codes and cross-border commercial litigation. Enforces cross-lingual consistency across regional language statutory circulars.
Capital Markets & Forensic Audits
Financial Institutions
Tabular Prospectus Context Financial Numeric Gate Corporate annual reports, audited financial prospectuses, and loan registries. Fast batch reranking, numerical table alignment, and footnote qualification verification.
Large-Scale Precedent Repositories
Apex Courts & Law Firms
Multi-Judgment Corpus Hierarchical Precedent Ranker Extensive regulatory repositories (>1,000,000 judgments) requiring interactive scoring. Balances expressive cross-encoder depth with sub-second retrieval across massive legal archives.

Complete Two-Stage Retrieval Workflow

How Sovereign SLMs, Bi-Encoders, and Precision Rerankers work together in the Harness Operating System.

End-to-End Orchestrated
Stage 01 • Query
Agent Statutory Request
An agent (e.g. §29A Screener) queries: "Has the PRA been classified as a willful defaulter?"
Stage 02 • Vector Recall
LexVector™ High Recall
Dense vector index returns 50 candidate passages from statutory circulars and NCLT orders in <10ms.
Stage 03 • Gating
PrecisionGate™ Cross-Encoder
PrecisionGate™ evaluates query vs each candidate, discarding false matches and ranking true hits (score > 0.85).
Stage 04 • SLM Reasoning
LexCore™ Output
The verified passages are injected into LexCore™ to generate the deterministic compliance report.