1
NPMI Normalized Pointwise Mutual Information
NPMI(i,j) = ln(P(i,j)/P(i)P(j)) / −ln(P(i,j))
Converts raw word co-occurrence into meaningful edge weights. Suppresses hub words ("the", "is"), amplifies rare but meaningful pairs.
2
GEE Graph Encoder Embedding
Z = Dα · A · Y
Model-free graph embedding. One matrix multiply — no eigen-decomposition, no gradient descent, no training.
3
LIDER Learned Index for Dense Retrieval
predict(position) → local binary search
Replaces ANN tree traversal with a learned model that predicts where each match lives. Sub-millisecond at billion scale.
Try:
🔬 Pipeline Trace
Step-by-step execution for this query
5 demo documents pre-loaded and indexed
Graph Theory · Machine Learning · Search Engines · NLP · Network Science
Try a suggested query above or type your own.
📄 Upload your own PDF (ephemeral — wiped on restart)
Drop a PDF here or browse
Max 20MB · replaces demo index until server restart🕸️ Co-occurrence Graph D3 force-directed network of word relationships
Nodes = vocabulary words (sized by frequency)
Edges = NPMI weight (thicker = stronger co-occurrence)
Click a node to see its nearest neighbors