Key Algorithms | Aevum Encyclopedia

Why Algorithmic Transparency Matters

Unlike black-box knowledge systems, Aevum publishes its core algorithmic architecture. Every module is peer-reviewed, version-controlled, and designed for maximum factual fidelity, minimal hallucination, and equitable cross-lingual performance.

Below you'll find our five foundational algorithms, their technical specifications, performance benchmarks, and how they interconnect to serve 2.4M+ verified articles in real-time.

🔍

Search & Retrieval

Vector Semantic Index (VSI)

Hybrid retrieval engine combining dense transformer embeddings with sparse lexical matching. Dynamically weights semantic vs. exact-match signals based on query intent classification.

94.2% NDCG@10

<12ms Avg. Latency

HNSW BM25+ Cross-Encoder Rerank Intent Router

🌐

Graph & Structure

Dynamic Knowledge Graph Builder (DKGB)

Continuous NER and relation extraction pipeline that maps entities, events, and concepts into a temporal graph. Uses GNN-based link prediction to resolve ambiguities and infer missing relations.

89.7% Link F1

4.1B Edges Indexed

GNN Temporal Reasoning Named Entity Disambiguation

🛡️

Verification

Cross-Reference Trust Network (CTN)

Multi-hop citation validation system that scores claim reliability using graph diffusion, source authority decay, and contradiction detection. Flags low-confidence assertions for human review.

99.1% False Positive Rate

0.03% Undetected Errors

PageRank Variant Contradiction Mining Trust Decay Functions

🌍

Multilingual

Alignment Engine (MAE)

Zero-shot cross-lingual mapping using contrastive learning on parallel and pivot corpora. Aligns concepts, citations, and metadata across 140+ languages with cultural context awareness.

91.4% Cross-Lingual BLEU

142 Supported Languages

M3M Contrastive Pivot Translation Contextual Embedding Space

📝

Curation & Output

Adaptive Curation Transformer (ACT)

RLHF-tuned abstractive summarizer and content router. Dynamically adjusts detail level, tone, and structure based on user expertise profile and query domain.

87.3% ROUGE-L

2.1x Readability Gain

RLHF Extractive+Abstractive Domain Adapter

How They Work Together

📥

Query Ingestion

Intent classification & language detection

→

🔍

VSI Retrieval

Dense + sparse candidate generation

→

🛡️

CTN Verification

Citation trust & contradiction check

→

🌐

DKGB Enrichment

Graph traversal & relation mapping

→

📝

ACT Output

Personalized synthesis & formatting

Core Retrieval Flow (Simplified)

                // VSI Hybrid Search Pipeline
                function executeQuery(query) {
                  const intent = classifyIntent(query);
                  const candidates = mergeResults(
                    bm25Search(query),
                    vectorSearch(encode(query), hnswIndex)
                  );
                  const verified = ctnFilter(candidates);
                  return actSynthesize(verified, intent);
                }
            

Open Research & Audit Access

All algorithm versions are published under CC-BY-SA 4.0. Peer reviews, benchmark datasets, and reproducibility scripts are available to registered researchers.

The Engine Behind Infinite Knowledge