NextPlaid is LightOn’s open-source multi-vector search engine, built for the kind of high-recall semantic retrieval that modern RAG systems and coding agents need. On top of it sits ColGREP, a semantic code search tool for your terminal and your agents. The project is released under the Apache-2.0 license, has about 599 stars on GitHub, and was actively updated in October 2026.
Multi-vector, or late-interaction, retrieval stores a vector per token instead of one vector per document, which preserves detail that single-vector indexes lose. NextPlaid packages that approach into a search engine usable from database workloads down to local tools. ColGREP shows the practical side: it combines classic regex filtering with semantic ranking, so you can narrow by pattern and still get meaning-based ordering. Everything runs locally and your code never leaves your machine. An optional agent mode downloads a small local model, about 2.5 GB, that explores a codebase like a sub-agent and returns the code locations that answer a question. Installation is available through Homebrew or a shell installer.
LightOn develops NextPlaid in the open on GitHub, and its documentation site covers both the ColGREP command-line tool and the NextPlaid engine itself, including the models used for multi-vector indexing of code and document collections.
NextPlaid and ColGREP are free and open source under the Apache-2.0 license. There is no cloud bill; the search and the optional agent model run on your own machine.
NextPlaid is for RAG builders who need stronger retrieval quality, and for developers who want fast, private, semantic search across large codebases from the terminal.