100% Private
No Signup
Free Forever
One of 64 free AI tools by Mahmoud Zalt.
Free Chat With Your Document (RAG)
A complete retrieval-augmented generation (RAG) pipeline that runs entirely in your browser, with no backend and no vector database. It splits your document into overlapping chunks, embeds each one with the all-MiniLM-L6-v2 sentence-transformer (Xenova/all-MiniLM-L6-v2) via Hugging Face Transformers.js, holds the vectors in memory, and uses cosine similarity to retrieve the passages most relevant to your question. Those passages become the context for a local language model (Llama 3.2 1B) served by WebLLM, the open-source high-performance in-browser LLM inference engine from mlc-ai, which streams a grounded, source-cited answer. Because both the embedding model and the language model load and run on your own machine, through WebAssembly for embeddings and WebGPU for generation, your document never leaves your device. No API keys, no accounts, no server.
Free and provided as is, without warranty. Use at your own risk. Terms
Free Semantic Search and RAG Without Sending Your Document Anywhere
Most "chat with your PDF" services upload your file to a remote server, store it, and run the AI in the cloud. That is a real privacy problem when the document is a contract, a medical record, an unpublished manuscript, or internal company material. This free tool takes the opposite approach: it runs the entire retrieval-augmented generation pipeline inside your browser. Your document is chunked, embedded, searched, and answered locally, so the text never travels across the network. The only download is the AI model files themselves, fetched once from a public CDN and then cached for instant reuse.
The workflow is simple. Load the embedding model and the language model, paste or upload your text, click Index, and start asking questions. Each answer is generated from the passages most relevant to your question and is accompanied by the exact source snippets and a similarity score, so you can always trace a claim back to the part of the document it came from.
How the WebLLM and Transformers.js RAG Pipeline Works
RAG combines two ideas: semantic search and language generation. First, the document is split into overlapping chunks of about 500 characters. Each chunk is converted into a vector embedding using all-MiniLM-L6-v2 (Xenova/all-MiniLM-L6-v2), a compact sentence-transformer run on-device with Hugging Face Transformers.js that captures meaning rather than surface words. When you ask a question, it is embedded the same way, and the tool ranks every chunk by cosine similarity to your question vector. The top matches become the context.
That context is handed to a local Llama 3.2 1B model served by WebLLM, described by its authors as a high-performance in-browser language model inference engine that leverages WebGPU for hardware acceleration, enabling powerful LLM operations directly within web browsers without server-side processing. WebLLM is the open-source @mlc-ai/web-llm project, is OpenAI-API compatible, and supports streaming chat completions, so answers appear token by token. The model is instructed to answer using only the supplied passages and to cite them, which keeps the response grounded in your actual document and reduces hallucination.
A Hands-On Way to Understand Private, In-Browser AI Search
Beyond being a practical document assistant, this tool is a working demonstration of the architecture behind modern AI search and enterprise knowledge assistants. The same pattern, embed, retrieve, then generate, powers internal documentation bots, customer-support copilots, and AI features inside note-taking and research apps. Seeing it run with no backend makes the moving parts concrete: you can watch indexing progress, inspect which chunks were retrieved, read the cosine-similarity scores, and judge how retrieval quality shapes the final answer.
For developers and teams evaluating RAG, this is a zero-setup sandbox built on proven open source. WebLLM (@mlc-ai/web-llm) and Hugging Face Transformers.js are both Apache 2.0 licensed, so the whole stack is auditable. There is no vector database to provision, no API budget to manage, and no data-handling agreement to sign, because everything happens on the user device. It is a fast way to build intuition before committing to a production stack, and a genuinely useful private tool in its own right for everyday document questions.
More Free Tools
More than 20 free AI tools.
Who asks documents questions instead of reading them cover to cover
Someone doing early due diligence on a contract or a vendor agreement pastes it in and asks pointed questions, what is the termination clause, what happens on late payment, rather than reading forty pages to find three answers. Students studying from a dense textbook chapter or a set of lecture notes ask conceptual questions and get an answer grounded in and cited to their actual source material, which is a meaningfully different (and more trustworthy) experience than asking a general chatbot that has no idea what edition or chapter they are studying from.
New employees onboarding into a role paste an internal handbook or a policy document and ask it directly rather than searching a wiki with imperfect keyword matching, and researchers querying a paper or a spec document use it to jump straight to the relevant section instead of skimming the whole thing looking for one detail.
How this compares to ChatPDF and NotebookLM
ChatPDF offers a free tier capped at a small number of pages and daily questions, with paid plans around 5 to 20 dollars a month for larger documents and higher limits, and every document is uploaded to their servers to be processed. Google's NotebookLM is free and genuinely capable, with strong citation behavior similar to what this tool aims for, but it requires a Google account and, like any cloud service, processes your document on Google's infrastructure rather than your own device.
This tool trades some of those services' polish and larger-model reasoning quality for complete locality: no account, no upload, no page cap tied to a subscription tier, because a 1B-parameter model running on your own GPU is simply a different scale of system than a data-center-hosted model. For confidential documents where the content itself cannot leave your machine, that tradeoff is the entire point; for a document where privacy is not a concern and you want the strongest possible reasoning over very long or complex material, a cloud service still has an edge.
Why retrieval quality matters more than model size here
A RAG system's answer quality depends less on how large the language model is and more on whether the right passages were retrieved in the first place, a small model given exactly the right three paragraphs will answer more accurately than a much larger model given the wrong ones. That is why this tool shows the retrieved snippets and similarity scores alongside every answer: if a response seems off or incomplete, checking whether the cited passages actually contain the relevant information is the fastest way to diagnose it, often the fix is rephrasing the question closer to the document's own terminology rather than blaming the model.
For the same reason, specific questions retrieve better than vague ones: asking "what is the notice period for termination" surfaces a tighter, more relevant set of chunks than asking "tell me about this contract," since the embedding model is matching your question's meaning against the document's chunks, and a specific question has a more distinct semantic signature to match against.
How It Works
Load the two models once: a small embedding model (about 30MB) that turns text into searchable vectors, and a local language model (about 1GB, WebGPU required) that writes the answers. Both cache in your browser after the first download.
Paste your text or upload a .txt or .md file, then click Index. The tool splits the document into roughly 500-character overlapping chunks and computes an embedding vector for each chunk in memory.
Ask a question. The tool embeds your question, ranks every chunk by cosine similarity, sends the top matches to the language model as context, and streams back an answer with the exact source snippets shown as citations.
Need expert help with AI?
Looking for a specialist to help integrate, optimize, or consult on AI systems? Book a one-on-one technical consultation with an experienced AI consultant to get tailored advice.
Key Features
Privacy & Trust
Use Cases
Limitations
- Answer generation requires WebGPU, which means a recent Chrome or Edge on desktop or laptop, or Safari 18+ on macOS. Embedding and retrieval still work without it through WebAssembly, but the WebLLM language model will not run.
- The local Llama 3.2 1B model is small and fast by design, so it excels at grounded question answering over retrieved passages but is less capable than large cloud models on complex multi-step reasoning or long-form writing.
- The first load downloads roughly 1GB for the Llama 3.2 1B weights plus around 30MB for the embedder. This is cached afterward, but the initial download needs a stable connection and some free disk space.
- Vectors are held in memory only, so the index clears when you reload the page or replace the document, and very large documents may be slow to index on low-end devices.
Frequently Asked Questions
Is this RAG tool free?
Yes, it is completely free with no signup, no API key, and no usage limits. There is nothing to pay because there is no server doing the work: the embedding model and the language model download once to your browser and then run entirely on your own device. It is built on free and open-source software, WebLLM (@mlc-ai/web-llm) and Hugging Face Transformers.js, both under the Apache 2.0 license, so you can use it as much as you like at no cost.
Is my document sent to a server?
No. Your document never leaves your device. Every step of the pipeline, splitting your text into chunks, embedding it into vectors, retrieving the most relevant passages, and generating the answer, runs locally inside your browser. The only network traffic is the one-time download of the AI model files from a public CDN, which then cache for reuse. You can verify this yourself by opening the Network tab in your browser DevTools while you index and ask questions: you will see no uploads of your document text.
Do I need to install anything?
No installation is required. This is a web tool that runs in your browser, with no extension, desktop app, or backend to set up. The first time you use it, the embedding model (about 30MB) and the Llama 3.2 1B language model (about 1GB) download in the background and cache in your browser, so subsequent uses are fast. You do not need to provision a vector database, sign up for an API, or configure any keys.
What is RAG and how does this tool use it?
RAG stands for retrieval-augmented generation. Instead of asking a language model to answer from memory, you first retrieve the most relevant pieces of your own document and feed them to the model as context. This tool does exactly that: it embeds your document into vectors with the all-MiniLM-L6-v2 model via Transformers.js, finds the passages closest in meaning to your question using cosine similarity, and asks a local Llama 3.2 1B model served by WebLLM to answer using only those passages. The result is an answer grounded in your text, with the source snippets shown so you can check them.
Why is the first load slow?
The first load downloads the two AI models that power the tool: the all-MiniLM-L6-v2 embedder, around 30MB, and the Llama 3.2 1B language model used by WebLLM, roughly 1GB. Downloading and initializing a billion-parameter model in the browser takes time, especially on a slower connection. Once downloaded, both models are cached by your browser, so every later visit and query starts almost instantly with no further download.
Do I really need WebGPU?
WebGPU is required for WebLLM to generate answers, because running a billion-parameter language model in the browser needs GPU acceleration. The embedding and retrieval steps run on the CPU through WebAssembly and work without WebGPU. If WebGPU is unavailable, the tool tells you and you can still index your document and inspect the retrieved passages, but you will need a WebGPU-capable browser such as recent Chrome, Edge, or Safari 18+ on macOS to generate the answers.
Does it work offline or on mobile?
After the first successful load, the models are cached, so the core pipeline can keep working without a live connection on the same browser. Mobile support is limited by WebGPU availability and device memory: generating answers with the 1GB Llama 3.2 1B model needs a capable, WebGPU-enabled device, which most phones cannot yet provide reliably. Embedding and retrieval are lighter and more broadly supported, but for the full chat experience a recent desktop or laptop browser is recommended.
What file types can I upload, and how large can the document be?
You can paste text directly or upload plain text files such as .txt and .md, up to roughly 200,000 characters, which covers a long article, a full contract, or a sizable report. For PDF or Word files, copy the text out and paste it in, or convert to plain text first. The document is split into roughly 500-character overlapping chunks and each chunk is embedded individually, so indexing time scales with length and device speed; very large documents take longer and use more memory.
Q&A SESSION
Got a quick technical question?
Skip the back-and-forth. Get a direct answer from an experienced engineer.