soulbah/regeste is an open-source project that runs a complete RAG pipeline in the browser. With 100% local parsing, embeddings, vector search, and chat, it offers a privacy-first alternative to cloud-based AI assistants.
The promise of conversational AI has a dark side: your data. When you upload a confidential document to a cloud LLM, you trust a third party with everything from legal contracts to medical records. Enter soulbah/regeste, an open-source project that takes a different approach. It combines Retrieval-Augmented Generation (RAG) and on-device AI to enable document chat entirely in the browser. Parsing, embeddings, vector search, chat, and citations never leave the user’s machine.
Regeste is a modern implementation of a local-first AI assistant for documents. The name may be unfamiliar, but the idea is simple: you upload a PDF or text file, ask questions about it, and get answers with citations—all without sending the document to a server.
The project is built with TypeScript and SvelteKit, showcasing a modern stack for privacy-preserving web applications. Its architecture is deliberately split: a browser-based data plane and a minimal server-side control plane. This design makes it a compelling reference for developers interested in edge AI and on-device inference.
Unlike a traditional cloud RAG system, Regeste does not require a hosted vector database, an external embedding API, or a remote LLM. The entire reasoning process happens on the user’s own hardware. That is a fundamental shift in how we can think about AI and data privacy.
Data privacy is not a niche concern. According to Grand View Research, the global edge AI market is projected to grow at a 36.7% CAGR between 2024 and 2030, reflecting rising demand for on-device intelligence. That demand is driven by a simple truth: many documents are too sensitive to upload.
As Edward Snowden wrote in Permanent Record, “Arguing that you don’t care about the right to privacy because you have nothing to hide is no different than saying you don’t care about free speech because you have nothing to say.” For businesses handling legal, medical, or financial data, privacy is a compliance requirement, not a preference.
Tim Cook echoed that sentiment at the IPAA Global Privacy Summit in 2018: “We believe privacy is a fundamental human right.” Regeste treats privacy as a technical requirement, not a marketing tagline.
RAG is a technique that combines retrieval of relevant document chunks with generative AI to produce grounded answers. Traditionally, RAG pipelines involve multiple servers: a document store, an embedding model, a vector database, and an LLM. Regeste does all of this in the browser.
When a user uploads a document, Regeste parses it locally. The parser extracts text and structures the content into chunks. Because everything runs in the browser, the original file and its parsed representation stay on the user’s device.
Next, the text chunks are converted into vector embeddings using an on-device model. These embeddings are stored in an in-browser vector store. When the user asks a question, Regeste encodes the query and runs a vector search locally, retrieving the most relevant chunks.
This is a significant departure from cloud-based RAG, where embeddings are generated on a server and stored in a hosted vector database. With Regeste, no network request is made for the core retrieval process.
The final step is answer generation. Regeste uses on-device inference to formulate a response based on the retrieved chunks. It also provides citations, showing exactly which parts of the document influenced the answer. This transparency is valuable for verification and trust.
Regeste is not entirely server-free. It uses a single Cloudflare Worker as its only backend component. This worker handles authentication, usage quotas, and the Assisted endpoint. It never touches document data. That means the server-side attack surface is minimal—fewer endpoints, less exposure.
The design choice is intentional. Edge infrastructure is used for control-plane concerns, while data-plane operations stay local. This split is a blueprint for other applications that want to minimize cloud dependencies without sacrificing features like user management.
The Assisted endpoint is the one optional exception to full local AI. It likely provides a helping hand for tasks that are difficult to run on-device, such as generating complex responses or improving accuracy. But even here, the user’s raw documents remain local, and the worker simply enhances the experience.
These features make Regeste stand out in a crowded field of AI chat tools. It is not just a theoretical concept; it is an operational implementation that developers can study, modify, and deploy.
Regeste targets three main audiences.
If you are creating an application that handles user documents, Regeste shows how to keep data on-device. Its architecture can serve as a reference for implementing local RAG pipelines. You can adapt its patterns for your own projects or contribute directly to the soulbah/regeste GitHub repository.
Law firms, hospitals, and financial institutions are bound by strict data protection regulations like HIPAA and GDPR. Regeste gives employees a way to use AI without violating compliance. Documents never leave the device, so there is no need to renegotiate vendor agreements or worry about data residency.
There is a growing audience of users who are wary of corporate data collection. For them, Regeste is a tool that respects their autonomy. They can ask questions of personal documents—medical letters, tax forms, contracts—without feeding them to a third-party model.
Cloud AI services are convenient and powerful, but they come with hidden costs. Ingesting a sensitive document into a public LLM can expose it to accidental leaks, training data usage, or subpoena. Even private cloud deployments require careful configuration and ongoing security audits.
Regeste’s browser-based approach eliminates many of these risks. There is no data transfer, no server-side document storage, and no external API call for the core RAG pipeline. The trade-off is performance: on-device models are often smaller and less capable than frontier cloud models. However, for many document comprehension tasks, that trade-off is acceptable—especially when the alternative is losing control of the data.
While Regeste is designed to be privacy-preserving, no system is foolproof. Users should be aware of the following best practices:
By following these practices, organizations can leverage Regeste without weakening their overall security posture.
The tech industry is shifting toward edge AI and on-device inference. The 36.7% CAGR projected by Grand View Research is a clear signal. Edge AI reduces latency, lowers bandwidth costs, and, most importantly, keeps sensitive data on the device.
Several factors are accelerating this trend:
Regeste is part of this movement. It is ahead of the curve in making full RAG pipelines feasible in a browser. As these technologies evolve, local AI applications will only become faster and more capable.
If you want to try Regeste, the project’s GitHub repository is the best starting point. After cloning the repository, you can set up the Cloudflare Worker for auth and quotas, then start the SvelteKit front-end. The documentation provides instructions for running the development environment and deploying a production version.
Contributions are welcome, particularly around improving the Assisted endpoint, adding support for more document formats, and optimizing on-device models. The project is open source, so you can inspect every line of code to verify its privacy claims.
While Regeste is a powerful demonstration, it has limitations. Browser-based inference can be slower than cloud-based AI, especially on older hardware. The quality of on-device models may not match large cloud models for complex reasoning. And the Assisted endpoint, while minimal, still introduces a server-side component—so it should be configured carefully.
The future of Regeste will likely be shaped by improvements in browser AI capabilities and the open-source community. As WebGPU and WebNN mature, performance will improve. As smaller, more capable models emerge, answer quality will rise.
Regeste proves that private document chat is possible without sacrificing the power of RAG. By running parsing, embeddings, vector search, and chat entirely in the browser, it eliminates the need to upload confidential documents to third-party services. The single Cloudflare Worker keeps the server-side exposure minimal, while TypeScript and SvelteKit deliver a clean modern experience.
If you are building privacy-sensitive AI tools or work with confidential documents, Regeste is worth a closer look. It offers a practical path toward a future where AI can be both intelligent and private. Explore the repository, test the approach, and consider how local-first AI can transform your applications.
soulbah/regeste is an open-source project that runs a complete RAG (Retrieval-Augmented Generation) pipeline entirely in the browser. You can upload a PDF or text file, ask questions about it, and get answers with citations—all without sending your document to a server.
Regeste performs parsing, embeddings, vector search, and chat inference locally on your own device. Because the data plane is browser-based and no external embedding API or remote LLM is used, your document content never leaves your machine. Only a minimal server-side control plane is needed for app orchestration, not for handling your data.
Local-first RAG means every step of the RAG pipeline—ingesting documents, generating embeddings, searching relevant passages, and running the language model—happens on-device. Traditional cloud RAG relies on hosted vector databases, remote embedding APIs, and cloud LLMs, which require sending sensitive data to third-party servers. Regeste eliminates that dependency in favor of edge AI.
The project primarily supports PDF and text files for uploading and chatting. Since the entire pipeline runs locally, your files are processed directly in the browser rather than being uploaded to a remote service. For the most current format support, check the repository's documentation.
The main limitations are related to hardware performance and model size. Running embeddings and chat inference entirely in the browser requires a reasonably capable device, and larger documents or complex questions may be slower compared to cloud-based models. However, the privacy and compliance benefits often outweigh these trade-offs for sensitive data.
