veles-core

Core library for Veles — fast, hybrid (BM25 + semantic) local code search for AI agents and humans, written in pure Rust.

veles-core is the indexing and search engine. It walks a directory, chunks source files, builds a BM25 inverted index plus a dense model2vec-rs embedding index, and serves hybrid queries with Reciprocal Rank Fusion. Tree-sitter is used to extract definitions for symbol-level lookups.

• No GPU, no transformer forward pass at query time. Static embeddings keep query latency in the tens of milliseconds on CPU.
• Persistent on-disk index under <repo>/.veles/, with incremental updates that reuse embeddings of unchanged files.
• Identifier-aware tokeniser that splits camelCase, snake_case, and mixed-script identifiers; multilingual model option for Cyrillic, CJK, Arabic, etc.
• Pure Rust — no Python, no protoc, no native ML runtime required.

Install

[dependencies]
veles-core = "0.2"

Quick start

use std::path::Path;
use veles_core::{SearchMode, VelesIndex};

# fn main() -> anyhow::Result<()> {
let index = VelesIndex::from_path(Path::new("."), None, None, false)?;

let results = index.search(
    "parse config file",
    5,
    SearchMode::Hybrid,
    None,  // alpha — auto-detect from query type
    None,  // language filter
    None,  // path filter
);

for r in results {
    println!("{} [{:.3}]", r.chunk.location(), r.score);
}
# Ok(())
# }

The first build downloads the default embedding model (~64 MB) into the HuggingFace cache (~/.cache/huggingface/hub/).

Persistence and incremental updates

use std::path::Path;
use veles_core::VelesIndex;

# fn main() -> anyhow::Result<()> {
let repo = Path::new(".");
let index = VelesIndex::from_path(repo, None, None, false)?;
index.save(repo)?;

// Later, reload without re-embedding:
let model = veles_core::model::load_model(None)?;
let mut reloaded = VelesIndex::load(repo, model)?;

// Files whose (size, mtime) fingerprint hasn't changed keep their
// embeddings; only modified/added files are re-embedded.
let report = reloaded.update_from_path(repo)?;
eprintln!("{} added, {} modified, {} removed",
    report.added_files, report.modified_files, report.removed_files);
# Ok(())
# }

What's in the box

See also

• The end-user CLI documentation and USAGE.md reference for the veles binary.
• veles-grpc — tonic-based gRPC service wrapping veles-core.
• veles-mcp — MCP server exposing search and find_related over JSON-RPC for AI agents.

License

MIT