What is snarevec?

SnareVec is a tool for capturing web pages you're reading and turning them into searchable vector data for your own self-hosted retrieval system, known as RAG. It has two parts: a browser extension that lets you click a button on any page to capture it, and a local background program called a daemon that does the actual work of extracting text, splitting it into chunks, generating embeddings, and storing them. This split exists because browser extensions are sandboxed and cannot run AI models, access arbitrary files, or connect to remote servers on their own, so the daemon handles everything that requires real system access. The daemon runs entirely on your own machine and listens only on the local network address, requiring a security token pasted from its startup message into the extension before anything can be sent to it. On first launch it sets up a working default configuration using a local embedding model, so pages can be captured immediately without further setup. Running the daemon automatically at login is optional and left entirely up to the user, with instructions provided for Windows, macOS, and Linux. Capturing a page happens through the extension's popup, where a profile is chosen, optional tags are added, and the page is sent through the extract, chunk, embed, and store pipeline. Failed captures are not lost, they land in a retry queue instead. The tool supports several different backends for generating embeddings, including a built-in local option, a local Ollama server, and any service exposing an OpenAI-compatible embeddings endpoint, along with several places to store the resulting vectors, including local files, a remote server over SSH, or directly into a Qdrant vector database server. A companion script allows vectors captured locally to be moved onto a remote machine running an embedded, single-process copy of Qdrant, with the destination service required to be fully stopped during the import and the script validating that the incoming data matches the destination's model and configuration before proceeding. The project enforces strict consistency rules, such as never mixing vectors from different embedding models within the same collection, since doing so would make search results meaningless.