Rust implementation of RAPTOR (Delling, Pajor, Werneck): given a public transit network, find all Pareto-optimal journeys between two stops, trading fewer transfers against earlier arrival.

Browser demo: https://urschrei.github.io/vulture/. WASM bindings (~290 KB gzipped) in vulture-wasm/, on npm as vulture-wasm.

The gtfs adapter implements Timetable over a parsed GTFS feed. aux/dmrc_gtfs.zip (Delhi Metro) is bundled.

use gtfs_structures::Gtfs;
use jiff::civil::date;
use vulture::{SecondOfDay, Timetable};
use vulture::gtfs::GtfsTimetable;

let gtfs = Gtfs::new("aux/dmrc_gtfs.zip")?;
let tt = GtfsTimetable::new(&gtfs, date(2024, 1, 15))?;

let start = tt.stop_idx("1").expect("Dilshad Garden");
let target = tt.stop_idx("44").expect("Vishwavidyalaya");

let journeys = tt
    .query()
    .from(start)
    .to(target)
    .max_transfers(10)
    .depart_at(SecondOfDay::hms(9, 0, 0))
    .run();

for j in &journeys {
    println!("{} trip(s), arrives {}", j.plan.len(), j.arrival());
}

Runnable: cargo run --example gtfs-timetable -- aux/dmrc_gtfs.zip 2024-01-15 1 44.

Each Journey carries plan: Vec<(RouteIdx, StopIdx)>, an origin, a target, and a label. journey.arrival() returns the SecondOfDay. For trip IDs and per-leg times, see Journey::with_timing.

Gtfs::new(path) reads from disk. gtfs-structures also exposes Gtfs::from_url, Gtfs::from_url_async, and Gtfs::from_reader; all yield a Gtfs that GtfsTimetable::new accepts.

let gtfs = Gtfs::from_url("https://example.org/feed/gtfs.zip")?;
let tt = GtfsTimetable::new(&gtfs, date(2026, 5, 4))?;

from_url requires the read-url feature on gtfs-structures. vulture pulls it with default-features = false; for the URL paths, depend on gtfs-structures directly: gtfs-structures = { version = "0.47", features = ["read-url"] }.

GtfsTimetable::station_stops(parent_id) expands a parent station into its child platforms. Passing the result to .from(...) / .to(...) runs a multi-source / multi-target search.

let tt = GtfsTimetable::new(&gtfs, date(2026, 5, 4))?;

let journeys = tt
    .query()
    .from(tt.station_stops("de:11000:900003201"))   // Hbf, ~301 platforms
    .to(tt.station_stops("de:11000:900100003"))     // Alex, ~50 platforms
    .max_transfers(10)
    .depart_at(SecondOfDay::hms(9, 0, 0))
    .run();

Berlin VBB (~42k stops, ~71k active trips): 385 µs.

HSL ships an empty transfers.txt. Build coordinate-derived walking edges before querying:

let tt = GtfsTimetable::new(&gtfs, date(2026, 5, 4))?
    .with_walking_footpaths(&gtfs, 500.0, 1.4);   // 500 m at 5 km/h

let journeys = tt
    .query()
    .from(tt.stop_idx("1040601").unwrap())   // Kamppi metro
    .to(tt.stop_idx("1453601").unwrap())     // Itäkeskus metro
    .max_transfers(10)
    .depart_at(SecondOfDay::hms(9, 0, 0))
    .run();

with_walking_footpaths builds bidirectional R-tree edges and preserves existing transfers.txt entries.

Paris IDFM is the largest of the bundled benchmark feeds (~54k stops, ~146k active trips per weekday). assert_footpaths_closed() opts into the single-pass O(E) footpath relaxation when transfers.txt is publisher-curated.

let tt = GtfsTimetable::new(&gtfs, date(2026, 5, 4))?
    .assert_footpaths_closed();

let journeys = tt
    .query()
    .from(tt.stop_idx("IDFM:monomodalStopPlace:45102").unwrap())   // Châtelet
    .to(tt.stop_idx("IDFM:monomodalStopPlace:44602").unwrap())     // Versailles RD
    .max_transfers(10)
    .depart_at(SecondOfDay::hms(9, 0, 0))
    .run();

Châtelet → Versailles RD: ~17 ms. Shorter cross-Paris queries (Châtelet → Gare du Nord, Châtelet → La Défense): under 1 ms. See docs/cross-city-benchmarks.md for the full table and methodology.

let profile = tt
    .query()
    .from(start)
    .to(target)
    .max_transfers(10)
    .depart_in_window(SecondOfDay::every(
        SecondOfDay::hms(17, 0, 0),
        SecondOfDay::hms(18, 0, 0),
        300,   // every 5 minutes
    ))
    .run();

Returns Vec<RangeJourney>: (depart, journey) pairs Pareto-filtered on (later depart, fewer transfers, earlier arrival). Serial .run() is rRAPTOR; .run_par() fans the per-departure work across a Rayon thread pool.

Allocate a RaptorCachePool once; each worker checks out a cache for the scope of one query.

use rayon::prelude::*;
use vulture::RaptorCachePool;

let pool = RaptorCachePool::for_timetable(&tt);

let results: Vec<_> = queries.par_iter().map(|q| {
    let mut cache = pool.checkout();
    tt.query()
        .from(q.from)
        .to(q.to)
        .max_transfers(10)
        .depart_at(q.depart)
        .run_with_cache(&mut cache)
}).collect();

pool.checkout() returns a PooledCache handle; the cache returns to the pool on drop.

Synthetic-network examples in vulture/examples/:

• cargo run --example quickstart – smallest viable RAPTOR query against a hand-built SimpleTimetable.
• cargo run --example reboarding – multi-route query where the recorded boarding stop is mid-route, not the route's first call.
• cargo run --example constraints – wheelchair, no-pickup, and no-drop-off boarding constraints.
• cargo run --release --example custom_label – custom Label with its own Ctx, route-preference scoring.
• cargo run --release --example fare_aware – fare-aware Pareto routing via ArrivalAndFare and a FareTable context.
• cargo run --release --example range_query – depart_in_window(...); serial rRAPTOR vs parallel naïve batch, with an equality assertion.
• cargo run --release --example cache_reuse – RaptorCache reuse with an equality assertion against one-shot results.

• Timetable trait – the network abstraction. GtfsTimetable covers GTFS; implement directly for non-GTFS data.
• Typestate query builder – tt.query().from(...).to(...).max_transfers(...).depart_at(...).run(). Order is compile-time enforced; .depart_at(...) / .depart_in_window(...) switch the return type.
• Multi-source / multi-target – .from(...) / .to(...) accept a StopIdx, a slice, a Vec, or (stop, walk_offset) pairs. Use station_stops(parent_id) for any-platform queries.
• Range queries – .depart_in_window(times) returns a Pareto profile. Serial path is rRAPTOR; .run_par() / .run_with_pool(&pool) use a parallel naïve batch.
• Label trait – default ArrivalTime for single-criterion. vulture::labels ships ArrivalAndWalk (arrival vs walking time) and ArrivalAndFare (arrival vs accumulated fare). Custom impls plug in via Query::with_context.
• Walking footpaths from coordinates – with_walking_footpaths(&gtfs, max_dist_m, speed_m_per_s) builds bidirectional R-tree edges, preserving existing transfers.txt.
• Closed-footpath fast path – assert_footpaths_closed() opts into single-pass O(E) relaxation instead of Dijkstra.
• RaptorCache – reusable scratch allocations per timetable. RaptorCachePool is the Sync variant.
• Per-leg timing – journey.with_timing(&tt, depart, origin_walk) attaches trip IDs and per-leg depart / arrive times to a Journey.
• Wheelchair-accessibility filter – require_wheelchair_accessible() skips trips and stops flagged NotAvailable in GTFS.
• Multi-day search – GtfsTimetable::new_with_overnight_days(&gtfs, date, OvernightDays(n)) loads n additional service days, shifting day-d trips by d × 86 400 seconds onto one time axis.
• Per-leg shape access – GtfsTimetable::shape_for_leg(&gtfs, trip_id, board, alight) returns the polyline points for one leg, sliced via shape_dist_traveled when present.
• Feed introspection – GtfsTimetable::features() returns a FeedFeatures snapshot. FeedFeatures::suggestions() returns an advisory string list mapping the snapshot to relevant vulture knobs.

Single-query latency. Warm RaptorCache; M-series Apple Silicon; single thread. Full table and methodology in docs/cross-city-benchmarks.md.

Range-query numbers (serial rRAPTOR vs parallel naïve batch): see the bench source and the linked benchmarks page. Head-to-head against the TypeScript raptor-journey-planner: docs/cross-impl-comparison.md.

The numbers above are measured with lto = true and codegen-units = 1 in this workspace's [profile.release]. Cargo does not propagate workspace profile settings to downstream crates; a vanilla cargo add vulture build inherits Cargo's defaults (lto = false, codegen-units = 16, opt-level = 3). To match these numbers from a downstream crate, set:

[profile.release]
lto = true
codegen-units = 1

docs/soundness.md audits the implementation issue-by-issue against the paper. The vulture-proptest harness runs the algorithm against a brute-force reference solver on every test invocation.

Three layers, all run by cargo nextest r on every PR:

• Unit tests in vulture/src/test.rs cover the algorithm (single-source, multi-source, range, footpath relaxation, wheelchair filtering, overnight) against hand-built SimpleTimetable fixtures and the bundled Delhi Metro feed.
• Doctests on every public type with a non-trivial contract (Label, Query, Journey, RaptorCache, GtfsTimetable::station_stops, etc.).
• Property-based tests in vulture-proptest generate random transit networks and check the algorithm against a brute-force reference solver, via Hegel.

The proptest harness has three generator layers (Layer 1: 1–2 routes, 2–4 stops, no footpaths. Layer 2: adds 1–4 footpaths. Layer 3: 1–4 routes, up to 6 stops, optional footpaths, multi-source / multi-target queries, per-route fares). Properties:

Hegel persists shrunk failing seeds to .hegel/. See vulture-proptest/README.md for the layer / issue map.

Run scripts/perf-smoke.sh before finalising any feature or Cargo.toml profile change. It runs the Delhi gtfs_query group from vulture/benches/gtfs.rs in criterion --quick mode against the last saved baseline in target/criterion/. Wall time: ~2 s with no source change, ~24 s after a source edit (the release profile's LTO link dominates). Criterion prints "Performance has regressed." when the median moves past noise.

The first run on a fresh checkout has no prior baseline; run twice, or pass --save-baseline <name> through to criterion (extra arguments are forwarded). Skip only when the change cannot affect runtime. The full criterion suite (cargo bench --features gtfs-bench, ~5 min) is the right thing to run before tagging a release.

For diagnosis after a regression: vulture/examples/profile-delhi.rs loops the same three queries with no criterion-harness overhead, sized for samply / cargo-flamegraph. Build with cargo build --profile profiling --example profile-delhi, then samply record -- target/profiling/examples/profile-delhi.

• parallel (default-on) – enables Query::run_par / Query::run_with_pool via rayon. default-features = false drops the dependency; RaptorCachePool stays available.
• gtfs-bench – the gtfs criterion benchmark.
• internal – the manual criterion benchmark over manual::SimpleTimetable.

Dual-licensed under either of:

• Apache License, Version 2.0
• Blue Oak Model License 1.0.0

at your option. SPDX expression: Apache-2.0 OR BlueOak-1.0.0.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this work shall be dual-licensed as above, without any additional terms or conditions.