greplm-core

Core indexing and search engine for greplm: a fast, offline trigram code index with code intelligence built for LLM agents.

The index lives in a .greplm/ directory at the project root. Immutable, mmap-backed segments combine trigram FSTs, roaring posting lists, and symbol tables. Search filters candidates by trigram intersection, then verifies matches with the real literal/regex matcher.

Features

Trigram content search — literal, regex, whole-word, with language and path filters
Symbol extraction — tree-sitter parsing for Rust, Python, JavaScript, TypeScript/TSX, Go, Java, C, C++, C#, Ruby, PHP, Swift, and Dart
Code intelligence — references, call graph (callers/callees/blast radius), typed go-to-definition, structural (AST) search
Context packs — task-driven, token-budgeted code assembly ranked by relevance and call-graph centrality
Git integration — blame, symbol history, changed-files-since-revision
Incremental indexing — only re-indexes what changed; optional file watcher and warm daemon
Fully offline — no network calls

Install

Add as a path or git dependency (the crate is not published separately on crates.io today; use the workspace or git URL):

[dependencies]
greplm-core = { path = "../greplm-core" }
# or
greplm-core = { git = "https://github.com/KhaledSMQ/greplm" }

Optional features:

greplm-core = { path = "../greplm-core", features = ["semantic"] }

Feature	Description
`semantic`	Offline semantic (vector) search with a trained Model2Vec embedder

Quick start

use greplm_core::search::{SearchQuery, SymbolQuery};
use greplm_core::Greplm;

fn main() -> greplm_core::Result<()> {
    let g = Greplm::discover(".")?;
    g.index(false)?;

    let searcher = g.searcher()?;

    let hits = searcher.search(&SearchQuery {
        pattern: "SegmentWriter".into(),
        limit: 10,
        ..Default::default()
    })?;

    let symbols = searcher.symbols(&SymbolQuery {
        name: "Searcher".into(),
        limit: 10,
        ..Default::default()
    })?;

    Ok(())
}

API overview

`Greplm`

The main entry point for a project and its index.

Method	Description
`Greplm::open(root)`	Open (and lazily initialize) the index for a project root
`Greplm::discover(start)`	Walk up from `start` to find the nearest `.greplm/` directory
`index(force)`	Build or refresh the index (incremental unless `force`)
`searcher()`	Open a `Searcher` over the current index
`status()`	Report segment count, document/symbol counts, last index time
`watch(debounce, on_change)`	Re-index incrementally on file changes
`compact()`	Merge segments and drop tombstoned documents
`clean()`	Remove the entire `.greplm/` directory

`Searcher`

Query interface over a loaded index. Key methods:

search — content search
symbols — symbol lookup (exact, prefix, substring, fuzzy)
references / references_resolved — text and structural references
callers / callees / blast_radius — call graph navigation
definition / references_of — typed go-to-definition at a position
structural_search — tree-sitter query or $NAME meta-variable patterns
context_pack — token-budgeted task context
outline / read_snippet — file structure and slices
blame / symbol_history / changed_since — git time-travel
summary — repository statistics

Other modules

Module	Purpose
`config`	`.greplm/config.toml` parsing
`lang`	Language detection and tree-sitter wiring
`client` / `daemon` / `proto`	Warm-index daemon over a Unix socket
`semantic`	Optional vector search (`semantic` feature)
`savings`	Token-efficiency tracking

Index layout

.greplm/
├── config.toml      # project configuration
├── meta.bin         # segment manifest and counts
├── cache/           # per-file hash cache for incremental indexing
├── segments/        # immutable mmap-backed index segments
└── savings.jsonl    # token-savings log (optional)

Related crates

greplm-cli — command-line interface (greplm)
greplm-mcp — MCP stdio server (greplm-mcp)

See the project docs for the full feature list, benchmarks, and agent setup.

License

MIT

greplm-core 0.6.0