Credits #

Pizza stands on the shoulders of the Rust and information-retrieval communities. This page credits the software we build on — and the projects whose ideas we learn from.

Reference points #

• Lucene — decades of query semantics and scoring design that the DSL and scoring model learn from.
• Elasticsearch — the JSON query DSL and REST API surface (_search, _doc, _bulk) the gateway is modeled on.
• tantivy — reference point for segment-organized inverted indexes and query execution.

Engine core #

The search engine (lib/engine) is a no_std-capable crate that compiles from server to browser; these crates make that possible:

• roaring — Roaring bitmaps carry every doc-id set: postings, filters, deletions.
• fst-no-std and fst-regex — our no_std forks of BurntSushi’s fst: finite-state term dictionaries, with regex and Levenshtein automata running directly over them.
• cedarwood — double-array tries for dictionary-driven analysis and suggestions (forked to drop its std dependency).
• hashbrown — the hash map that works everywhere, including the wasm builds.
• memmap2 — immutable segments are memory-mapped, never copied.
• bitpacking — SIMD-packed posting lists.
• zstd (with lz4 and flate2) — segment, column, and stored-source compression; a single .fire segment stacks all three where they win.
• rayon and crossbeam — parallel query execution across segments and workers.
• rhai — scripting for scripted aggregations and pipelines.
• turbovec — TurboQuant 4-bit vector quantization for approximate kNN.
• Apache Arrow (parquet, arrow-array, arrow-schema) — mounting Parquet files directly as searchable segments.
• moka — segment cache, ureq — remote segment fetching, and fs4 — advisory file locks for the KV store.
• nom — parser combinators behind the query_string mini-language.
• wildmatch — wildcard query and route matching.
• crc32fast and crc32c — CRC checksums over segments, WAL frames, and manifests; corruption is detected, never guessed around.

Distributed layer #

• openraft — the Raft consensus that keeps cluster state and the catalog in agreement; we vendor it with openraft-rt-monoio, a runtime binding that lets Raft’s timers and heartbeats run on our monoio reactor instead of tokio.
• capnp (with capnp-rpc) — the inter-node transport wire format; zero-copy both ways.
• monoio — the thread-per-core, io_uring-based runtime the gateway and replication build on, with signalfut handling OS signals inside that runtime.
• tokio — the async runtime for the pieces whose ecosystems assume it: the capnp-rpc transport endpoints and the actix-based modules. The split is deliberate — monoio owns the data path, tokio sits at the edges.
• rskafka — the client behind the Kafka WAL backend, via our monoio-compatible fork.
• sled — the embedded KV store holding Raft log and catalog state on disk.

Formal methods #

• TLA+ / TLC — the protocols we built on top of Raft are specified and model-checked in contrib/formal-models before being trusted in code: promotion and WAL handover, shard allocation, online rolling reshard, doc-id reservation, inplace-column seeding, layered deletion visibility, and the membership fence. The TLC counterexamples have caught 18+ real bugs along the way.

Gateway and operations #

• actix-web — HTTP types and client plumbing in the gateway.
• serde and serde_json — every byte of the JSON surface, schema included.
• clap — the CLI.
• tracing — structured logs and the per-request trace tree.
• jemallocator — the global allocator, with profiling and stats enabled.
• sysinfo — host and disk facts behind node stats.
• flatten-json-object — flattening nested JSON documents into dot-pathed fields.
• arc-swap — lock-free swapping of routing tables and security snapshots: readers never pause while a new cluster state publishes.

Text analysis #

The per-language analyzer family (contrib/analysis-*, 40+ languages) is a Lucene-lineage port with embedded dictionaries, leaning on the best Unicode infrastructure available:

• ICU4X — Unicode segmentation, normalization, casing, and collation.
• jieba-rs — Chinese segmentation, alongside the SmartCN, IK, simplified↔traditional (stconvert), and pinyin analyzers.
• lindera — Japanese morphological analysis with the ipadic dictionary (the kuromoji analyzer), and Korean via nori (ko-dic).
• Stempel and Morfologik — Polish stemming and morphology, straight out of the Lucene contrib family.
• The European stemmer set (analysis-english, -french, -german, …) and the Indic, Thai, Turkish, and Vietnamese analyzers round out the family.

Bindings and this site #

• pyo3 — native Python bindings for the engine (lib/engine/bindings/python): the same segments and query engine, importable from Python.
• meilisearch-docsearch — the search box in these docs is pizza-searchbox, our fork of the docsearch component that the Meilisearch and Tauri ecosystems use.

Browser builds #

• wasm-bindgen (with js-sys and web-sys) — the engine compiled to WebAssembly, the same segment format mounted read-only in the docs playground.
• serde-wasm-bindgen — zero-copy search results from wasm into JS.

Maintainers and contributors are listed in the repository and on the community page. Thank you all.