lightonai/next-plaid

NextPlaid, ColGREP: Multi-vector search, from database to coding agents.

What it solves

NextPlaid and ColGREP address the limitations of standard vector search, which collapses entire documents into a single embedding. This "lossy" approach often fails for complex data like source code, where critical details (parameters, docstrings, control flow) are lost. NextPlaid provides a high-performance multi-vector database to preserve this detail, while ColGREP applies this technology to create a local, semantic code search tool for developers and AI agents.

How it works

NextPlaid uses a multi-vector approach (similar to ColBERT) where each document is represented by multiple embeddings rather than one. To make this scalable, it employs product quantization (2-bit or 4-bit compression) and memory-mapped indices to keep the RAM footprint low. It includes built-in ONNX Runtime for server-side encoding and uses SQLite for metadata pre-filtering.

ColGREP leverages NextPlaid to index codebases. It uses Tree-sitter to parse code into structured representations (extracting signatures, calls, and docstrings) before embedding them with a ColBERT-style model. This allows users to perform semantic searches combined with regex filtering, all running locally on the CPU.

Who it’s for

  • Developers who want a fast, local, and privacy-preserving way to search their codebases semantically.
  • AI Agent creators who need to provide coding agents (like Claude Code) with high-precision retrieval tools.
  • ML Engineers looking for a general-purpose, local-first multi-vector database for retrieval workloads.

Highlights

  • Local-first execution: ColGREP is a single Rust binary; code never leaves the machine.
  • Multi-vector precision: Maintains multiple embeddings per document to avoid information loss.
  • Hybrid search: Combines semantic ranking with regex filtering.
  • Efficient storage: Uses product quantization and memory-mapping to fit millions of documents in memory.
  • Agent-ready: Built-in installation hooks for agents like Claude Code, OpenCode, and Codex.
  • CPU-optimized: Designed for high performance on CPUs, with optional CUDA support.

Related

  • Project
    kantord/SeaGOAT

    A local code search engine that uses vector embeddings and ripgrep to enable semantic and keyword-based search across a codebase.

  • Project
    PromtEngineer/localGPT

    A private, on-premise document intelligence platform that uses hybrid search and local LLMs to let users query and summarize their files without data leaving their machine.

  • Project
    MinishLab/semble

    Semble is a CPU‑only Python library/CLI/MCP server that indexes a codebase in ~0.5 s and answers natural‑language code queries in ~1 ms, delivering transformer‑level relevance while using ~99 % fewer tokens than a grep‑plus‑read baseline. It combines static code embeddings (Model2Vec) with BM25, fuses results, and applies code‑aware reranking. Integrates with Claude Code, Codex, Cursor, OpenCode, etc., and requires no API keys or GPUs.

  • Project
    infiniflow/infinity

    An AI-native database designed for LLM applications that provides high-speed hybrid search across dense and sparse embeddings, tensors, and full-text data.

  • Project
    illuin-tech/colpali

    ColPali is a research library for visual document retrieval. It encodes document images with vision‑language models (e.g., PaliGemma, Qwen‑VL) into multi‑vector embeddings and scores them against query embeddings using a ColBERT‑style late‑interaction method. The original `colpali-engine` package (now deprecated) provides model classes, processors, optional fused kernels, fast‑Plaid indexing, token‑pooling, and interpretability tools. New projects should use Sentence‑Transformers v6’s `MultiVectorEncoder`, which integrates ColPali‑style models.