Traceable by design
Component ranks, matched terms, retrieval methods, latency, and experiment identifiers are recorded instead of hidden.
A research system, not a black box
Scientific discovery should be explainable, testable, and honest about uncertainty.
PaperMetrix is a research platform for retrieving and recommending scientific papers. It combines established information-retrieval methods with modern scholarly embeddings, while preserving the evidence needed to understand and evaluate every ranking.
RRFTransparent fusion
BM25Lexical
S2Semantic
GraphCitation
- Papers
- 0
- Authors
- 0
- Scholarly sources
- 0
- Open-access records
- 0
Our philosophy
Useful recommendations require visible evidence.
A ranking can be helpful without pretending to be an objective measure of scientific quality.
Time-aware evaluation
Temporal benchmarks prevent a system from recommending papers that did not exist at the simulated query date.
Humans remain the judge
Offline metrics and interaction logs support analysis, but independent relevance judgments remain essential for publishable claims.
How it works
From fragmented metadata to an inspectable ranking.
- 01
Acquire
OpenAlex, Semantic Scholar, Crossref, arXiv, and Unpaywall contribute complementary metadata and access signals.
- 02
Normalize
DOIs and source identifiers are reconciled into a canonical scholarly record with provenance and missing-data reasons.
- 03
Retrieve
BM25 captures exact terminology; SPECTER2 captures scholarly meaning; graph signals can contribute citation structure.
- 04
Fuse and explain
Reciprocal Rank Fusion combines positions without pretending that unrelated score scales are directly comparable.
- 05
Evaluate
Frozen splits, request-level telemetry, independent judgments, and reproducible artifacts connect results to defensible claims.
What is new here
LLMs are an auditable channel—not the foundation of truth.
PaperMetrix introduces generative models only where they can be isolated, disabled, priced, and compared against a fixed baseline.
Query expansion is versioned by model and prompt, protected by budgets and circuit breakers, and currently restricted to staff experiments.
Lexical, semantic, and graph channels remain independently observable before fusion.
Temporal integrity, frozen development/test partitions, pooling, agreement, and cost/latency reporting are first-class parts of the platform.
Impressions, clicks, saves, and relevance feedback can be linked to the exact retrieval request without turning behavior into automatic truth.
A clear boundary
What PaperMetrix does—and does not do.
It helps you
- find related and topically relevant literature
- inspect why a result was retrieved
- organize papers into reusable research collections
- run reproducible retrieval experiments
It does not claim to
- measure scientific quality from rank alone
- replace systematic-review screening
- treat citations or clicks as unbiased relevance
- let an LLM silently rewrite the benchmark
Follow the evidence trail from query to paper.
Search the current corpus and inspect the retrieval signals behind every live result.
Open scholarly search →