rag visualizer — opening
Loading
A high-precision developer tool and visual walkthrough that exposes the mathematical mechanics normally concealed between a query and an LLM response — then executes the entire pipeline for real on your private documents directly inside your browser.
Drop any document into Your document to run real token chunking, 384-dimension embeddings, and interactive 3D Semantic Vector Space. Switch to Guided trace for an instant, pedagogical walkthrough with verified ground truths.
ASP.NET Core runtime, data access, and service architecture
Pipeline Trace
/Question
Question: Start a run to watch the query move from representation to retrieved evidence and answer.
A guided walkthrough of a Retrieval-Augmented Generation pipeline. Pick a knowledge domain, select a question, and watch it get embedded, matched against document chunks, assembled into a grounded prompt, and answered. This mode is deterministic — the scores and answers are authored so the run is always coherent. Switch to Your document to run the same pipeline for real on your own file in this browser.
Guided trace is an authored benchmark with deterministic 3D orbit coordinates and full citation provenance. Live mode computes real in-browser embeddings and cosine similarities on any file you upload. Neither mode simulates what can be computed.
| Pipeline Stage | Guided Trace Benchmark | Your Document (Live WebGPU) |
|---|---|---|
| Chunking | One chunk per curated document scenario. | Sentence-aware split, ~800 chars with 15% overlap, sized to the embedding model's 256-token context window. |
| Embedding | Seeded deterministic projections with 3D spherical orbit & cosine distance radar visualization. | all-MiniLM-L6-v2 (23 MB, quantized) in dedicated WebWorker — 384 real dimensions, mean-pooled and normalized. |
| Retrieval | Interactive 3D Semantic Vector Space with orbital camera, Top-K cutoff sphere, and 2D radar projection. | Exact dot-product cosine similarity against all indexed chunk vectors, projected into 3D space with citation linking. |
| Context | Segmented token budget gauge (System + Query + Evidence) with bidirectional citation linking. | Isolated <evidence> XML boundary envelope with delimiter escaping to neutralize indirect prompt injection. |
| Generation | Authored ground-truth response with interactive citation pills linked to source chunks. | SmolLM2-360M-Instruct executing on WebGPU with WASM fallback, or 1-click clipboard prompt handoff. |
Designed & Engineered by Abhishek Suman
Building resilient backend architectures with observability interfaces that bring clarity to complex AI systems.