RPG

Quick Start

Node Types

Every node in the graph is classified as one of three types:

Edges connect nodes to show relationships — for example, a Method that addresses a Research Question and produces a Finding.

Features

RPG Similarity Algorithm

The Related Papers feature ranks every library paper against the current paper using a weighted Jaccard similarity score computed over shared taxonomy categories of Q, M, and F nodes. The pipeline has four stages:

Stage 1 — Distinctive vocabulary construction

The taxonomy contains ~252 L2 sub-categories (≈ 10 per node type × ~8 sub-categories × 3 types) each with a name, description, and hundreds of representative example sentences. For each category, a raw vocabulary is assembled from the union of all tokens in its name, description, and example sentences.

Because plant biology papers share many generic terms (e.g. gene, expression, protein), a raw vocabulary would cause every category to match every paper. A cross-category IDF filter is applied: a token is kept only if it appears in ≤ 25 % of the L1 categories for that node type. This retains the terms that are distinctive to each category while discarding domain-wide noise.

Stage 2 — Node classification (top-K per node)

For each Q, M, or F node label in a paper, asymmetric coverage is computed against every category's distinctive vocabulary:

Asymmetric coverage is used rather than symmetric Jaccard because category vocabularies are much larger than short node labels; symmetric Jaccard would penalise exact nodes unfairly. The top-2 L1 and top-2 L2 categories with score > 0 are assigned to each node. Each paper's final category sets are the union of its per-node assignments, capturing all research themes covered by the paper.

Stage 3 — Weighted Jaccard similarity

Given two papers A and B, each with six category sets (Q-L1, Q-L2, F-L1, F-L2, M-L1, M-L2), the similarity score is:

where Jaccard(X, Y) = |X ∩ Y| / |X ∪ Y|.

Default weights (user-adjustable via sliders):

L2 weights equal the corresponding L1 weight multiplied by the L2 specificity bonus slider (default 2×). All six weights are normalised to sum to 1 before computing the final score.

Stage 4 — Ranking & caching

All 2,600+ papers are scored against the query paper (≈ 10–30 ms per request). Papers with a zero score (no shared category at all) are excluded. Results are paginated (8 per page) and the score distribution is displayed as a percentage in each result card (green ≥ 40 %, blue ≥ 20 %, grey < 20 %). When using the default weights, results are cached in memory (LRU, 200-entry cap) for fast repeat access. The category index is built once at server startup and cached to disk (_similarity_index.json), so subsequent restarts are nearly instant.

BM25 Node Search Algorithm

The Node Search feature ranks individual Q, M, and F node labels from all The Plant Cell papers in response to a free-text query. It uses BM25 (Best Match 25), a classical information-retrieval algorithm that is highly effective for scientific text because authors use precise, domain-specific terminology where lexical overlap is a reliable proxy for semantic similarity — without requiring a large language model or GPU.

Corpus & index

Approximately 150,000 node labels (Q, M, F) are indexed as individual documents. The index is built once at server startup and cached to disk (_node_search_index.json) so subsequent restarts load in under 2 seconds. Each node stores its paper provenance (paper ID, title, year) alongside the label and its token list.

Tokenisation

Node labels are lowercased, stripped of punctuation, and filtered with an 83-word stopword list (generic English words plus common scientific verbs such as demonstrate, identify, analyze). Unlike the similarity classifier, BM25 keeps the token list (not a set) so that term frequency is preserved.

BM25 scoring

For a query Q and document D:

IDF down-weights terms common across many node labels; tf-normalisation rewards relevant term repetition while penalising very long labels. Parameters k₁ = 1.5 and b = 0.75 are standard BM25 defaults that work well across diverse corpora.

Score normalisation & result modes

Within each result set the top-scoring node is assigned 100 % and all others are scaled proportionally, making scores interpretable across different queries. Three filled dots (•••) indicate > 66 %, two dots > 33 %, and one dot any positive score.

In All mode the top 5 results per type are retrieved in a single request and scored independently (so Q top = 100 %, M top = 100 %, etc.). Switching to a specific type activates paginated mode (10 results per page). Typical query latency is under 15 ms for 150,000 nodes.

Tips & Shortcuts

FAQ