About
TryRAG is a free online RAG playground for teaching, learning and prototyping retrieval-augmented generation (RAG). It exists for one reason: these ideas are much easier to understand when you can watch them run on a document you know.
Why it was made
Most explanations of RAG are either diagrams with boxes and arrows or code that needs an account, an API key and an afternoon to set up. Neither lets a beginner see what actually happens to their text. Here you paste a document, ask a question and see every intermediate result: the chunks, the numbers each chunk becomes, the similarity sum, and the prompt that would be sent to a model.
The tools and what each is for
- Try RAG on your own document: Learn how RAG works by watching every step run on a real document: splitting, vectors, similarity scores, retrieval and the final prompt.
- Evaluate RAG retrieval: Measure how well retrieval works on your own document, with standard scores and how far each could move.
- Chunking Visualizer: See where a document will be cut before you build a RAG system, and spot chunks that break mid-sentence or mix topics.
- Token Counter & Cost Estimator: Learn what tokens are and how they relate to words and characters.
- Compare RAG Setups: Check that an updated document still answers the questions the old one did, before you replace it.
- Compare document parsers: See whether one parser keeps the information your questions need, such as table rows with their names, headings and numbers.
- Evaluate your RAG output: Score a RAG system that already exists, from its logs or a test run, without rebuilding it here.
- Debug a wrong RAG answer: Find out why a RAG system gave a wrong answer to one question: missing data, chunking, retrieval or the model.
- Graph RAG Playground: Learn what a knowledge graph is by seeing one built from a document you know.
- RAG Cost Calculator: Estimate the monthly cost of a RAG system before building it.
Every feature, grouped by what it is for: all tools.
The inference playgrounds
A second set of pages covers what happens after retrieval, when a model on a GPU reads the prompt and writes the answer. They are listed on the Inference page.
- Prefill vs Decode: Estimate the time to first token and the tokens per second of a model on a GPU before renting it.
- KV Cache Calculator: Check whether a model, a context length and a batch size fit on a GPU before renting it.
- Quantization Visualizer: See what quantization does to individual weights, not only to the file size.
- Batching Simulator: See why a server that waits for the longest request in a batch wastes most of its capacity.
- Prefix Caching Simulator: See how much of each prompt a prefix cache can skip, and the first-token time it saves.
- Speculative Decoding Simulator: Understand how speculative decoding produces several tokens from one pass of the large model.
- LLM Inference Cost Calculator: Work out what a self-hosted model costs per million tokens at a given utilization.
The guides explain the ideas behind each tool in plain language.
Who it is for
- Students and people new to AI who want to understand RAG, embeddings, chunking and tokens by trying them.
- Teachers and trainers who need a live demonstration that works in any browser, with nothing to install.
- Developers who want to check how a document splits, what would be retrieved and what a prompt would cost before writing code.
How it works
Everything runs in your browser. The core engine is compiled to WebAssembly. An optional small language model can be downloaded once to search by meaning, and you can bring your own model (BYOM), such as Ollama, Gemini or Claude, to get written answers. There is no account, and no server of ours processes your text.
What it is not
It is a place to learn and experiment with small documents. It is not a hosted RAG service, it does not store documents, and programs cannot connect to it. Token counts and costs are estimates for learning and rough planning.
Questions, corrections or suggestions are welcome: get in touch.