indexkit

Offline hybrid retrieval: build a local semantic + lexical index over your files and query it. Chunks and embeds a corpus through ollama, indexes it with turbovec (a quantized vector index), and fuses that with SQLite FTS5/BM25 for opt-in hybrid search.

No API key and no network egress by default — the embedding endpoint is localhost. A configured remote CONTEXT_KIT_OLLAMA_HOST receives corpus chunks and queries, so that is the one setting that sends data off the machine.

Notes-first but corpus-agnostic: loaders are pluggable, so the same engine can index Markdown notes, code, or any text corpus.

Formerly local-rag. The old name claimed a deployment property that a supported setting falsifies, and understated an engine that also does lexical retrieval. Pre-rename environment variables are still honored.

Install

pip install indexkit          # or: uv tool install indexkit
ollama pull nomic-embed-text  # once

That is the whole setup — no plugin host, no bootstrap step. Indexes default to ${XDG_DATA_HOME:-~/.local/share}/indexkit.

indexkit index ~/notes --name notes
indexkit query "how did we handle retry backoff" --name notes --k 8

To work from a clone instead — for contributing, or to run an unreleased revision:

# From a clone of https://github.com/mbeacom/context-kit
pip install ./plugins/indexkit    # or: uv tool install ./plugins/indexkit

As a context-kit plugin

Claude Code and GitHub Copilot CLI auto-bootstrap the bundled bin/indexkit launcher from the plugin's SessionStart hook, into ${CLAUDE_PLUGIN_DATA}/venv — for Copilot, ~/.copilot/plugin-data/<marketplace>/indexkit/venv. In that mode the host controls where indexes live.

The launcher prefers that bootstrapped venv, then falls back to an indexkit already on your PATH, then to any importable indexkit module. So a pip-installed copy satisfies the plugin too, and a missing venv is not fatal.

Requirements

• ollama running with an embedding model pulled:

  ollama pull nomic-embed-text
  

• uv — only for the plugin bootstrap path. A pip install needs nothing beyond Python 3.10+.

For APM or manual plugin usage — or on any host where the venv is missing or stale — bootstrap it yourself into a neutral data location:

export CONTEXT_KIT_DATA="$HOME/.local/share/context-kit"
bash scripts/bootstrap.sh
export PATH="$PWD/bin:$PATH"

To check readiness without installing anything:

bash scripts/bootstrap.sh --check   # exit 0 ready, 3 needs bootstrap

It prints KEY=VALUE lines: status (ready, missing, stale, uv-missing), the raw venv_status and uv availability as separate fields, the resolved home/venv paths, and the exact bootstrap_command. A stale venv was built from different pyproject.toml metadata and would run outdated code, so it is reported as clearly as a missing one. Because uv is only needed to build the venv, an already-usable venv reports ready even when uv is absent. Dependent tooling uses this to detect an unusable runtime on hosts that do not deploy plugin hooks, such as APM. (Claude Code and GitHub Copilot CLI both run this plugin's SessionStart bootstrap.)

Usage

Index a corpus, then query it:

indexkit index <path> --name X
indexkit query "your question" --name X
indexkit query "exact terms and intent" --name X --hybrid
indexkit status --name X
indexkit list
indexkit remove --name X --yes

Each named index is persisted under <data-dir>/indexes/<name>/, so queries are fast and survive across sessions. The data directory resolves in this order:

1. CONTEXT_KIT_DATA (or the PRODUCTIVITY_SKILLS_DATA alias)
2. CLAUDE_PLUGIN_DATA, set by a plugin host
3. ${XDG_DATA_HOME:-~/.local/share}/indexkit — the standalone default

An existing ~/.claude/plugins/data/indexkit directory still wins over the standalone default while that default has not been created, so upgrading a plugin install does not orphan indexes you already built.

Index lifecycle

Index names remain backward-compatible with earlier releases: any non-empty single path component except . or .. is accepted, including names with spaces or more than 80 characters. Path separators (/ and \) and NUL are rejected. These containment rules apply consistently to index, query, status, and remove.

indexkit remove --name X --yes permanently removes one named index. The command is non-interactive and refuses to run without --yes; missing indexes fail clearly. Indexing, querying, status inspection, and removal share a per-index process lock, so removal fails clearly while that index is in use. Once locked, removal moves only the selected index out of the active namespace, then unlinks its flat artifact files without recursive directory deletion. Other indexes are untouched, and incomplete cleanup is reported with the quarantined artifact location rather than silently ignored.

Portable environment variables:

None of these are required for a standalone install; all have defaults. The pre-rename CONTEXT_KIT_LOCAL_RAG_HOME is still read as a fallback for CONTEXT_KIT_INDEXKIT_HOME.

CONTEXT_KIT_DATA normally holds both the venv and the indexes. Set CONTEXT_KIT_INDEXKIT_HOME only when a caller needs to redirect index data to an isolated store while still using one shared bootstrapped venv — the memory plugin does this to keep each project's index inside its own project-isolated provider directory. When it is unset, behavior is unchanged.

The pre-rename PRODUCTIVITY_SKILLS_* names still resolve as a deprecated alias (CONTEXT_KIT_* → PRODUCTIVITY_SKILLS_* → Claude fallback).

Hybrid retrieval

Semantic-only retrieval remains the default. --hybrid adds SQLite FTS5 lexical BM25 candidates and fuses them with turbovec semantic candidates using deterministic reciprocal-rank fusion: 1.0 / (60 + semantic_rank) + 1.0 / (60 + lexical_rank). Each source retrieves 3 × k candidates before fusion, so the candidate depth is greater than the requested final result count. JSON results include source offsets and per-source rank/score metadata; text output remains compact.

indexkit query also accepts an --allowlist of candidate documents (read from a file, or - for stdin), which applies to both semantic and lexical candidates. For example, feeding Obsidian backlinks into a hybrid query:

obsidian backlinks file="X" | indexkit query "..." --hybrid --allowlist -

FTS5 is detected and backfilled automatically for existing indexes. indexkit status reports its fts5 capability. If the SQLite build lacks FTS5, semantic retrieval continues to work but --hybrid exits with a clear error.

MIT © Mark Beacom.