Menu
Request Demo
Home / How it works
The semantic pipeline

The semantic pipeline
behind every reviewer
match.

Three stages – concept extraction, semantic search, and reviewer ranking. Working together to identify the most relevant reviewers for any manuscript in under 60 seconds.

01 Stage 01 AI-powered concept extraction Title + abstract → subject areas + ranked concepts 02 Stage 02 Semantic reviewer search Research concepts matched through Milvus-powered search 03 Stage 03 Three-tier ranking Concept → Similarity → Coverage, ranked by relevance
— Stage 01

AI-powered concept
extraction

An AI model trained on academic literature analyzes the manuscript’s title, abstract, and keywords to generate structured outputs that power the reviewer search.

Output A
Subject areas

Broad academic disciplines (e.g. "Computational Biology," "Machine Learning"). These drive Subject Area Coverage scoring.

Output B
Ranked concepts

Specific research ideas scored 0.00–1.00. A weight of 0.92 means central; 0.31 means peripheral. This drives the Concept Score.

Review extracted concepts before running the search.

Reviewer match quality depends on accurate concept extraction. If a key concept is missing, you can refine the abstract or add a targeted keyword and run the extraction again.

extraction.complete
2.4s
Subject areas
Computational Biology Machine Learning Genomics
Weighted concepts
0.00 — 1.00
transformer architectures
0.92
attention mechanism
0.88
protein folding prediction
0.74
graph neural networks
0.58
stochastic gradient descent
0.31
benchmark datasets
0.22
14 concepts · editor-approved
FIG. 01 — Editor view: extracted concepts ranked by weight
— Stage 02

Vector database search
powered by Milvus

Extracted concepts and subject areas are converted into vector embeddings and queried against an indexed article database using Milvus. This is fundamentally different from keyword search.

Legacy
approach.legacy

Keyword search

Matches exact strings
Misses synonyms, evolving terminology, and cross-disciplinary connections. Biased toward well-known terms.

Query "attention mechanism"
"attention mechanism" survey match
"transformer architectures" no string match
"self-attention layers" no string match
Reviewer Select
approach.semantic

Semantic vector search

Matches meaning
A reviewer writing about "transformer architectures" surfaces for a paper on "attention mechanisms" even with no keyword overlap.

Query vector [0.84, -0.12, …, 0.31]
"transformer architectures" cos 0.91
"self-attention layers" cos 0.87
"encoder-decoder framework" cos 0.79
— Stage 03

Reviewer aggregation and
three-tier ranking

Potential reviewers are extracted from matched articles. Three scores are computed and combined into a final rank using a strict priority order.

Primary · Tier 1 01

Concept score

Semantic similarity between your extracted concepts and the reviewer's article concepts. The most important signal.

0.90
concept · cosine
Secondary · Tier 2 02

Similarity score

Overall semantic similarity combining subject areas and concepts. Provides a holistic alignment view.

0.75
similarity · holistic
Tertiary · Tier 3 03

Concept coverage

The extent of your manuscript's concepts semantically covered by this reviewer's work. The breadth signal.

0.60
coverage · fraction
Strict priority order

Concept → Similarity → Coverage

Concept Tier 1 Similarity Tier 2 Coverage Tier 3
Ranking reflects semantic expertise alignment only.

Institutional prestige, h-index, career stage, and network proximity do not influence rank order they appear as filterable informational metrics only.

h-index · informational institutional rank · informational career stage · informational network proximity · informational

Watch how the pipeline runs on
your own manuscript.

Paste an abstract get a ranked, evidence-backed reviewer shortlist in under 60 seconds.

Request a demo Back to overview