Everything you can do on TryRAG
Every tool and feature on the site, grouped by what it is for. Everything here is free, needs no account and runs in your browser. New to RAG? Follow the learning path.
Retrieval, step by step
- Ask your own documentsPaste text, load a web page or upload a PDF, text, Markdown, CSV, JSON or HTML file. Ask a question and see the chunks a RAG system would retrieve and the prompt it would send.
- Every step shownSplitting, embeddings, the similarity sum, a map of the chunks and the final prompt, each opened up with the real numbers.
- Several sources at onceSearch up to eight documents together, see which source each chunk came from, or limit the search to one.
- Word search or meaning searchMatch by exact words, or turn on a small language model in your browser to match by meaning.
Better retrieval
- Hybrid searchCombine word and meaning search by a blend of scores or by rank fusion (RRF), and see where each chunk stood in both lists.
- Re-rankingScore the best candidates a second time with a cross-encoder in your browser or with your own model, and see which chunks moved.
- Evaluate with test questionsWrite questions with the words the answer must contain and get hit rate, MRR, precision, recall and nDCG, with ranges and a downloadable report.
- Auto-tuneTry chunking, overlap, search method, top-k and re-ranking on your questions and get a recommended setup, with the reason for each choice.
- Debug a wrong answerPaste your document, the question, the wrong answer and the right one, and see whether data, chunking, retrieval or the model went wrong.
- Evaluate your own RAG outputUpload what your own system retrieved and answered, as CSV or JSON in Ragas columns, and get retrieval and answer scores and a comparison of two runs.
- Why was it not found?Name the text you expected and get the reason it was missed, with setting changes that are checked before they are suggested.
Beyond the basics
- Chunking visualizerSee exactly where a document is cut by fixed size, sentence or heading, with overlap highlighted, and export the chunks.
- RAG interview preparationCommon AI engineer interview questions on retrieval, evaluation and serving, each with a short answer and a tool to try it on.
- Compare RAG setupsRun the same questions against two document versions, up to three sets of settings or two prompts, and see what got better or worse.
- Compare parsersLoad a document and the text another parser extracted from it, and see which reading lets your questions find the right passage.
- Graph RAGBuild a small knowledge graph from a document and see which extra chunks following its links brings in.
- Bring your own modelConnect Ollama, Gemini, Claude or any OpenAI-compatible model for written answers, embeddings and re-ranking.
Cost
- Token counterEstimate the tokens in a prompt and what it costs per request and per month on popular models.
- RAG cost calculatorEstimate what a RAG system costs to build and run, compare two setups and export the working.
Questions
Can I use it for prototyping a RAG system?
Yes. Load your real documents, try chunk sizes, overlap and matching settings, measure which works best on your own questions, and compare two setups side by side. The settings you find carry over to any RAG framework you build with afterwards.
Can I try RAG online without installing anything?
Yes. The whole tool runs in your browser. Paste text or upload a PDF, ask a question and you see the retrieved chunks straight away. There is no account, no install and nothing is uploaded.
Does it have hybrid search and re-ranking?
Yes. Choose Hybrid under the document and the tool combines word matching and meaning matching, by a blend of scores or by rank fusion (RRF), and shows where each chunk stood in both lists. Re-rank scores the best candidates a second time, with a small model that runs in your browser or with your own model, and shows which chunks moved up or down.
Can I use a PDF and a web page together?
Yes. Press Add another source and load each one: a PDF, a text file, a link or pasted text, up to eight. One question searches all of them, each retrieved chunk shows which source it came from, and you can limit the search to a single source.
Does TryRAG support BYOM (bring your own model)?
Yes. Open "Bring your own model" and connect a model you already use: Ollama on your own computer, or Gemini, Claude or any OpenAI-compatible service with your own API key. The model then writes the answers. Requests go straight from your browser to the model, the key is not saved, and nothing passes through this site.
How do I connect my own model?
Open "Bring your own model" under the document and enter the address and model name. Pick the Ollama, Gemini or Claude example to fill in the address, then add your model name and key. Local programs such as Ollama and LM Studio work if they accept OpenAI-style requests. You may need to allow this site in the program's settings. Requests go straight from your browser to the address you enter.
How do I evaluate and tune retrieval?
In the Evaluate tab, add questions and the words the right chunk must contain. "Evaluate current settings" shows how many questions find that chunk. "Auto-tune" tries chunking, overlap, search method, top-k, diversity and, when they are on, re-ranking and the minimum score, one step at a time. It recommends the cheapest setup that is statistically as good as the best, and says why.
What is the difference between matching by words and by meaning?
Word matching finds text that shares words with your question and works instantly. Meaning search uses a small language model that runs in your browser, so it also finds text that says the same thing in different words. It needs a one-time download of about 25 MB.
Can my program connect to this tool?
No. It runs only on this page, for learning and small documents. There is no server or API to connect to.