ix

Sub-millisecond code search via sparse trigram indexing.

ix builds a compressed trigram index that is typically 2-3× the source size for pure code, and can be smaller than the source for repetitive or binary-heavy repos (measured: 0.13× on a 1 GB mixed-content repo). The compaction pipeline — delta encoding → protobuf varint → ZSTD level 3 — achieves 88% reduction vs raw u32 storage and 60% additional savings on top of varint alone. The CDX trigram table uses a B-tree page architecture (block index → ZSTD-compressed 1024-entry blocks) for sub-50μs random access into compressed data.

This eliminates the linear-scan bottleneck of traditional tools on large codebases. Target hardware floor: 2015 CPU, 8 GB RAM.

Documentation

Install

cargo install moeix

Installs two binaries:

• ix — CLI search tool
• ixd — background daemon (requires notify feature, enabled by default)

You only need ix for search. Install ixd if you want continuous indexing.

Quick Start

# Build the index
ix --build /path/to/repo

# Literal search
ix "fn validate"

# Regex search
ix --regex "fn\s+\w+_handler"

# Context lines around each match
ix --context 3 "TODO"

# Show query statistics
ix --stats "struct Config"

# Only matching file paths
ix --files-only "error"

# Count matches only
ix --count "TODO"

# Filter by file extension
ix --type rs --type py "fn main"

Daemon

ixd watches one or more directories for file changes and incrementally updates the index. It builds the base index on startup, so you do not need to run ix --build before it — running both rebuilds the same shard twice (verified: ix --build ran in ~12 ms on a 40-file / 240-trigram fixture, then ixd spent ~11 s rebuilding the identical base; source: src/lib/daemon.rs:232 unconditionally calls builder.build()). Use ix --build only for a one-shot, no-daemon index.

# Foreground (debugging, or wrapped by systemd)
ixd /path/to/repo

# Detach into the background and return immediately (v0.13.0+)
ixd --daemon /path/to/repo

# Watch multiple directories (v0.9+)
ixd /project-a /project-b /project-c

Expected startup output (v0.13.1, captured 2026-08-13):

ixd [repo-name]: watching ...
ixd [repo-name]: loaded config — 3 exclude patterns, 0 watch roots
ixd [repo-name]: initial build complete (40 files, 240 trigrams)
ixd [repo-name]: socket at /run/user/1001/ixd/{hash}.sock

Each directory runs on its own thread with independent index, watcher, beacon, and Unix domain socket. Signal handling and memory monitoring are shared.

Service Management (Linux / systemd)

# Install as a user-level systemd service
ix service install /path/to/repo

# Start / stop / restart / status the service
ix service start
ix service stop
ix service restart
ix service status

The service auto-starts on login and survives reboots. See docs/DAEMON-RUNBOOK.md for full operation guide.

Daemon Socket

The daemon exposes a Unix domain socket for external consumers (editors, tooling):

$XDG_RUNTIME_DIR/ixd/{hash}.sock

Protocol is NDJSON — one JSON object per newline-terminated line. See docs/SOCKET-API.md. The ix CLI reads the index file directly, not through the socket.

Index Statistics

Show what's inside an index — file count, trigram count, on-disk size:

# Human-readable
ix stats /path/to/repo

# JSON output
ix stats --json /path/to/repo

This is distinct from ix --stats "pattern" (the search flag that shows per-query statistics). ix stats inspects the index itself.

Configuring the Daemon

Scope what ixd watches and ix --build indexes with .ixd.toml:

# .ixd.toml
watch_roots = ["src", "lib"]
exclude_patterns = [".git", "node_modules", "target", "vendor"]

See docs/.ixd.toml.md for full schema and examples.

How It Works

1. Extract — ix --build walks the directory, extracts byte-level trigrams (skipping null bytes to nullify binary noise), and caps at 64 offset samples per trigram for files >1 MB.
2. Accumulate — Trigrams are grouped into posting lists (one per unique trigram). An external sort with 500K-entry flush threshold keeps RAM constant regardless of repository size.
3. Compress — Posting lists and the trigram table use the same pipeline: delta-encode adjacent file IDs and offsets → protobuf varint → ZSTD level 3. The CDX trigram table is organized as a B-tree: a 12-byte-per-1024-entry block index for O(log N) lookup, then decompress one ~5 KB block to find the target.
4. Plan — On search, the query is decomposed into trigrams. The block index finds the target block, one ZSTD call decompresses it, and a linear scan finds the posting list offset.
5. Verify — Candidates are filtered through per-file bloom filters (256 B, 0.7% false-positive rate), then streamed through a regex matcher with constant memory usage.

Compaction Pipeline (measured)

Raw u32 entries  →  delta-encode  →  varint  →  ZSTD level 3
  10.6 MB             2-3× smaller     60% more     88% total reduction
                                                      (1.3 MB final)

Index Format (v1.3)

All integers little-endian, offsets absolute from file start, 8-byte aligned.

CDX compression is always-on since v1.3. Not backward compatible with v1.1/v1.2 — rebuild indexes after upgrading:

rm -rf .ix/
ix --build .

Performance

Measured on a 2015-era CPU (Haswell equivalent), 8 GB RAM. All ratios verified from actual indexes.

ix wins when the trigram index eliminates most files from scanning. On small repos or queries where every file matches, linear-scan tools like ripgrep are faster.

Feature Flags

Library

ix is also a library (moeix on crates.io, ix as the crate name):

[dependencies]
moeix = "0.12"

use ix::reader::Reader;
use ix::executor::{Executor, QueryOptions};
use ix::planner::Planner;

let reader = Reader::open(".ix/shard.ix")?;
let plan = Planner::plan("struct Config", false);
let mut executor = Executor::new(&reader);
let (matches, stats) = executor.execute(&plan, &QueryOptions::default())?;

See docs.rs/moeix for the full API reference.

Building

cargo build --all-features
cargo test --all-features
cargo clippy --all-features -- -D warnings

Requires Rust 1.85+.

License

MIT