Article · Microsoft Learn
How retrieval-augmented generation works Opens in a new tabLearn how documents are chunked, embedded, retrieved, and used to ground an answer.
Loading indexed vectors and preparing their map.
Explore the concepts, research, and open-source tools used in RAG Playground.
Learn how retrieval, embeddings, vector search, and generation work together.
Article · Microsoft Learn
How retrieval-augmented generation works Opens in a new tabLearn how documents are chunked, embedded, retrieved, and used to ground an answer.
Video · IBM Technology
What are Word Embeddings? Opens in a new tabLearn how word embeddings turn language into vectors that capture meaning and similarity.
Video · IBM Technology
What is a Vector Database? Powering Semantic Search & AI Applications Opens in a new tabLearn how vector databases store embeddings and retrieve similar content.
Video · IBM Technology
What is retrieval-augmented generation? Opens in a new tabLearn how retrieval and generation combine to produce grounded answers.
Paper · arXiv
The original RAG research paper Opens in a new tabRead the foundational research introducing the RAG architecture.
Learn how embeddings become explorable maps and node-link diagrams show relationships.
Article · UMAP documentation
How UMAP works Opens in a new tabSee how UMAP turns high-dimensional embeddings into a map you can explore.
Interactive · PAIR
Understanding UMAP Opens in a new tabExplore UMAP concepts and tradeoffs through interactive visual explanations.
Paper · arXiv
UMAP for dimension reduction Opens in a new tabRead the original paper describing UMAP's theory, algorithm, and applications.
Guide · Visualizing.org
How to read node-link diagrams Opens in a new tabLearn how nodes, links, clusters, hubs, and bridges reveal structure in a network.
Learn about the open-source tools used to build RAG Playground.
Learning resources open in a new browser tab.
Indexing, embeddings, chunking, and stored source controls.
Indexed sources and storage in this browser.
Embedding model
all-MiniLM L6 v2
Indexed items
0
A sentence-transformer model that embeds each whole chunk for similarity search and retrieval.
Xenova/all-MiniLM-L6-v2Category
Sentence transformer
Cache status
Not downloaded
First download
~24 MB
Model files stay in this browser's cache and can be evicted when storage is constrained.
No embedding models cached in this browser.
Saved with My project for future indexing. Each stored source separately keeps the settings that created its current vectors.
Every embedding model can index text. These settings control how source text is divided before embedding.
Keeps nearby sentences together until the chunk reaches a character budget.
Set how much sentence-level context fits into each stored chunk before the next chunk begins.
See which resources belong to this project and which are shared across projects in this browser.
My project
Sources, indexed items, and saved layouts are isolated from other projects.
0/500 indexed items
Completed layouts retained for this project and recreated when needed.
0 also retained in this session for this project
These resources are shared across projects in this browser profile.
Origin storage
Usage unavailable used
Unavailable capacity · Unavailable
Projects
0
Cached models
0
Active runtimes
0
Storage estimates include every project, shared app data, and any other data owned by this origin.
Embedding · Corpus-fitted lexical
local/tfidf-v1
Embedding · Static word
NeuML/fasttext-quantized
Embedding · Sentence transformer
Xenova/all-MiniLM-L6-v2
Embedding · Image-text
Xenova/clip-vit-base-patch32
Embedding · Audio-text
Xenova/clap-htsat-unfused
Embedding · Text-image-audio
antflydb/clipclap
Generation · Instruction model
onnx-community/Qwen3-0.6B-ONNX
Generation · Vision-language model
HuggingFaceTB/SmolVLM-500M-Instruct
Interaction · Hand tracking
https://storage.googleapis.com/mediapipe-models/hand_landmarker/hand_landmarker/float16/1/hand_landmarker.task
Models download from their owning workflow when first needed. Removing cached files keeps indexed content intact and preserves assets shared with another cached model.
No resources active or retained in memory
0 working · 7 stopped
Loads sentence-transformer models and embeds indexing or query text.
Starts automatically when needed.
Loads quantized word vectors and embeds indexing or query text.
Starts automatically when needed.
Loads the local language model and generates grounded answers.
Starts automatically when needed.
Reduces indexed-item vectors into the two- or three-dimensional map.
Starts automatically when needed.
Builds map links, layout coordinates, metrics, and semantic groups.
Starts automatically when needed.
Streams project exports and validates and stages project imports.
Starts automatically when needed.
Loads the hand-landmark model and processes camera frames.
Starts automatically when needed.
Choose whether to keep downloaded models. Both actions remove every project, cached results, and local settings, then reload the website.
Browser permissions and the persistent-storage policy may remain after the reset.
Browser storage sizes are origin-wide estimates. Exact per-database and per-worker memory usage is not available consistently across browsers.
0 sources · 0 indexed items
Add files, import a IIIF manifest, search Wikipedia and public media, or paste source text. Content is split into indexed items, embedded, and stored locally.
Drop files here or choose files
Up to 20 files · 100 MB total
Capture from your phone
Pair a phone with a short-lived encrypted session.
Presentation API 3.0 and 2.x · text, images, and audio only
Open a separate live dictation experiment using this computer or a paired phone.
Open Voice word map0 sources · 0 indexed items
Retrieve semantically similar indexed items. Text queries can also generate a local answer from supported results.
Model: all-MiniLM L6 v2 · Text only
Query with
Enter to retrieve · Shift+Enter for a new line
Download an embedding model to enable retrieval.
0 sources · 0 indexed items
Discover ways to inspect, navigate, analyze, and interact with the map. Hover or focus a guide item to locate its interface, then select it to reveal the related control.
The graph panels remain available in Upload and Query. Close this guide whenever you want the full canvas.