Montre

A modern, embeddable query engine for corpus linguistics. Lightweight, easy to install and use, but powerful.

montre (/mɔ̃tʁ/): to show; to reveal; to make visible (French)
From Latin monstrare “to point out, indicate.”

Montre is particularly suited for aligned literary corpora and multi-edition corpora.

Status

Early development. Not yet usable for real research work. The architecture and data model are stabilizing; APIs, formats, and CLI are still in flux.

What is Montre?

Montre is a local-first corpus query engine:

• No server
• No daemon
• No global registry
• No service dependencies

A Montre corpus is a portable artifact: a single directory containing indexed text, annotations, and (optionally) alignments. You can open it from the CLI, Python, Julia, or R -- as a library, not as a service.

Goals

• Fast queries on large annotated corpora (100M+ tokens)

• Embeddable: use as a library, not a server

• Native NLP integration:

  • CoNLL-U
  • Stanza JSON
  • UDPipe
  • spaCy exports

• Clean, expressive query language based on CQL

• First-class parallel corpus support:

  • multiple languages
  • multiple editions
  • multiple competing alignments
  • alignment-aware querying

Design principles

1. Corpus as artifact

A corpus is a build product, not a runtime configuration:

• immutable
• reproducible
• portable
• versionable

This enables:

• stable research artifacts
• reproducible experiments
• reliable citation

2. Components of a corpus

A corpus may contain multiple components:

• monolingual subcorpora
• reference corpora
• editions
• translations

Each component is independently queryable, but can participate in structured relations (alignments).

3. Alignments as data

Parallelism of corpora is flexible.

Alignments are:

• named
• typed
• layered
• replaceable

You can have several alignments (over sentences, paragraphs, etc) potentially from different models (e.g. LaBSE, vecalign) and choose which alignment(s) to use at query time.

4. Alignment-native querying

Queries can project across alignments:

<lemma="bibelot"] =labse_sentence=> component:"maupassant-en"

This is projection, not a join:

• hit sets move between components
• cardinality may change
• relations are explicit and named

---

Capabilities (planned)

Querying

• token queries
• span queries
• structural queries
• metadata filters
• distributional queries
• alignment projection

Parallel corpus support

• sentence-level alignments
• paragraph-level alignments
• many-to-many mappings
• multiple competing alignment models
• edition-aware alignment

Interfaces

• CLI
• embeddable Rust API
• Python bindings
• Julia bindings
• R bindings
• TUI (ratatui-based, separate repo)

---

Building

cargo build --release

---

Usage (planned)

# Build a corpus from CoNLL-U
montre build --input corpus.conllu --output ./my-corpus

# Query
montre query ./my-corpus '[pos="NOUN"] [pos="NOUN"]'

# Info
montre info ./my-corpus

---

Library usage (planned)

Python

import montre

corpus = montre.open("./my-corpus")
for hit in corpus.query('[pos="DET"] [pos="NOUN"]'):
    print(hit.start, hit.end)

Julia

using Montre

corpus = open_corpus("./my-corpus")
for hit in query(corpus, "[pos=\"DET\"] [pos=\"NOUN\"]")
    println(hit)
end

---

Architecture

montre-core     Core data model (Position, Span, Token, Unit, Component)
montre-index    Index structures (inverted, forward, span indexes)
montre-query    Query parser, planner, optimizer, executor
montre-build    Corpus construction (CoNLL-U, JSON, text + metadata)
montre-align    Alignment ingestion and projection engine
montre-cli      Command-line interface
montre-py       Python bindings
montre-jl       Julia bindings

---

License

Apache-2.0