Skip to content

Demo 02 · Client-side

Keyword vs semantic retrieval

Type a query — or load a preset — and watch the same corpus ranked two ways: BM25 keyword scoring on one side, a concept-map proxy for semantic similarity on the other. Some queries only one side gets right; that is the point. The best-ranked passages are then stitched into a cited answer.

The corpus covers my own projects — read the case studies →

Both rankings are computed live in your browser. The semantic column is a lexical proxy — token overlap over a hand-built concept map, not an embedding model — and it is labeled that way wherever a score appears.

Presets

Semantic wins — the best passage describes the speedup without ever saying "faster".Keyword wins — a rare exact term appears verbatim in exactly one passage.Semantic wins — a paraphrase that shares almost no tokens with its target passage.Keyword wins — the semantic proxy over-generalizes "fast" and drifts to the wrong project.Neither side answers — the corpus has nothing on this, and the floor refuses.

Semantic wins — the best passage describes the speedup without ever saying "faster".

designed cases: semantic wins 2 · keyword wins 2 · neither answers 1

Runs locally in your browser.

What’s in the corpus — 12 passages

CodLab

  • Author disambiguation at scale
  • The O(N²) read pattern
  • Out of core, one machine
  • Signals, vetoes, and safeguards

Mitsubishi

  • YOLO inference optimization
  • Both axes at once

DIATICS

  • Full-cycle computer vision
  • Plate recognition and pose estimation
  • Serving models with FastAPI

Senior project

  • An assistant for Alzheimer's patients
  • Recall in conversation
  • Medication tracking and exercises