vaultspec-rag 0.2.25

GPU-accelerated semantic search for your docs and source code: hybrid dense + sparse RAG on a local Qdrant backend, served to Claude Code and other MCP clients.

vaultspec-rag: semantic search for a vaultspec-core workspace

A vaultspec-core project accumulates a durable record of decisions, plans, research, and the code they produced. vaultspec-rag searches that record and your source code by meaning, not by keyword.

Search "file lock concurrent write per-root" and vaultspec-rag surfaces the decision that governs it, even when the document never uses those exact words. It is the retrieval layer of the project: it finds and ranks the grounding, and a client such as an AI assistant reads it.

The architecture overview explains how it works; the glossary defines the terms used across the docs.

Requirements

Before you install, confirm your machine meets these minimum requirements:

• Python 3.13 or newer
• uv as the package manager
• An NVIDIA GPU with CUDA support
• About 3 GB of free GPU memory
• Linux or Windows

macOS, AMD GPUs, and Apple Silicon are not supported. The architecture overview explains why the hardware floor sits where it does.

Quickstart

Install

Add vaultspec-rag to your project and set it up:

uv add vaultspec-rag
uv run vaultspec-rag install
uv sync

install configures the GPU PyTorch build, downloads the search models, and provisions the managed search server. uv sync then pulls in that GPU build. The models total a few gigabytes, so the first download takes several minutes, but it runs only once.

Index and search

1. Start the server:

   uv run vaultspec-rag server start
   

2. Index your project:

   uv run vaultspec-rag index
   

3. Search:

   uv run vaultspec-rag search "concept plus the domain terms"
   

The first run builds the index. After that, the running service watches your files and reindexes changes automatically, so the index stays current without another command. See the getting started guide for the full walkthrough.

Searching by meaning

The index is hybrid. A semantic half matches concepts and a keyword half matches exact terms, so write your query as a short phrase that both describes the concept and names the domain terms the target text would use. Pure prose starves the keyword half.

uv run vaultspec-rag search "reentrant collection lock ordering during indexing"

1. adr/2026-06-12-service-concurrency-adr
   adr | feature: service-concurrency | status: accepted | 2026-06-12
   Store-layer locking distinguishes local mode (one reentrant lock per
   collection) from server mode (no point locks). The lifecycle lock is
   always acquired before any collection lock.
2. reference/2026-06-12-service-concurrency-reference
   reference | feature: service-concurrency | 2026-06-12
   Per-root locks keep concurrent writers from colliding on the shared store.

Each result is a rank, a location you can open, and the matching text. Vault hits carry a metadata line, so a superseded ADR shows as superseded before you read it.

Searching code and filtering

Search code with --type code, and narrow with filters including --language, --path, and a symbol name. Add --scores to see the relevance number beside each rank:

uv run vaultspec-rag search "gpu section wrapping the reranker predict forward pass" --type code --language python --scores

1. src/vaultspec_rag/search/_searcher.py:308 (score 0.9241)
   # The GPU lock wraps only the model forward call; the
   # score-to-float conversion below runs after release.
   with self._gpu_section(timings):
       while True:
           try:
               raw_scores = reranker.predict(

For the full filter set (path globs, document type, feature, date), see search and index.

Documentation

Getting started

• Getting started - install, index, and run your first query end to end.
• Installation - the GPU build, dependency provisioning, and recovery steps.

Daily use

• Search and index - run searches and refresh the index.
• Service mode - keep models warm in a background service for faster queries.
• Backends - the managed Qdrant server versus local-only mode.
• MCP integration - wire search into Claude Code and other MCP clients.
• Automation - JSON output and scripting.
• Preprocessing hooks - index PDFs, spreadsheets, and other formats.

Reference

• CLI reference - every command and flag.
• Configuration - settings, environment variables, and defaults.
• Service discovery - the service.json contract for integrators.
• Glossary - terms used across the docs.

Concepts

• Architecture - how it works, why a GPU is required, and the server and local-only modes.
• Indexing - indexing and retrieval internals.

Support and help

File bugs and ask questions on the GitHub issue tracker.

A good bug report carries five things: your vaultspec-rag version, your operating system, your GPU model, the exact command you ran, and the full stderr output. With those, a maintainer can reproduce the fault. Without them, the report is hard to act on.

Changelog and license

The changelog holds release notes and version history.

vaultspec-rag is released under the MIT License. See LICENSE for the full text.