Context Engineering

Token Optimization & Matryoshka Embeddings

An interactive benchmark exploring context compression and cost reduction on LLM inputs using Matryoshka Representation Learning (MRL) and semantic retrieval filters.

Gemini APIMatryoshka EmbeddingsRAG PipelineTypeScript

Code Specs

Bundle Size 14.2 KB

Est. LOC 840 lines

Deps 4 npm pkgs

Languages

TypeScript 75%

Javascript 15%

CSS 10%

Experiment Objective

To evaluate and prove budget optimization when retrieving semantics rather than feeding full document context into Gemini models.

Telemetry & Metrics

38% Max Cost Savings

83% Vector Dimension Compression

~99% Cosine Sim Retention

Live Simulator / Console

token-optimization-embeddings_preview.sh

$ node run_rag_simulator.js

✓ Loaded gemini-embedding-2 (Dimensions: 768)

ℹ Vector database initialized with 4 document chunks.

$ query --text "How much is the security deposit?"

★

Launch Simulator

Interactive terminal rendering dynamic vector states, debug feedback, and remote sandbox telemetry under resource isolation.

Environment & Runtime

Host Platform Google Cloud Run

Orchestrator Node.js Express

Engine / Language TypeScript & Gemini SDK

Context / Isolation Vite Environment Sandbox

Key Features

⚡ Vector dimensionality scaling from 1536, 768 down to 256 dimensions using Matryoshka Representation Learning (MRL).
⚡ Dual-model generation evaluation using gemini-3.5-flash and gemini-3.1-pro-preview.
⚡ Real-time comparison between full context (Naive) and retrieved chunks (RAG) metrics.
⚡ Optimized costs ($0.001140 vs $0.001824 USD on 256 dim) with minimal impact on semantic cosine retrieval (0.6735 max similarity).

Core Hypothesis & Outcomes

Proves that reducing vector dimensions from 1536 (Pro) to 256 (MRL Ahorro) using Gemini's native embedding models retains high cosine similarity while delivering up to 38% cost savings and 83% compression in index size.