S
SUJAN
Information Retrieval · EXPERIMENTING

RAG Visualizer

Inspecting retrieval pipelines, token fragmentation, and vector distance distributions in real-time.

TELEMETRY: 0.42s · 256 chunks · 89% precision
TOOLS: TypeScript Canvas FastAPI Cosine Sim

Hypothesis & Experiment

Visualizing high-dimensional embedding clusters in 2D projection helps engineers identify semantic blind spots and chunk boundary fragmentation prior to model ingestion.

Findings

  • Visual inspection quickly spotted duplicate chunks caused by improper markdown header splitting.
  • Visualizing cosine distance distributions revealed that threshold filtering at 0.72 optimal cutoff discarded 80% of noise without missing true citations.