Switchify Prediction

Offline English word completion using a Rust library and bundled SQLite. The read-only baseline and a separate personal database feed an in-memory n-gram predictor. No keystroke capture, background monitoring, abbreviations or Switchify PC integration is included.

The default model is en-aac-oanc-v1: 21,674 words and 539,151 n-grams. It adds AAC-focused communication and a controlled spoken-English sample to the earlier WorldAlphabets/Taskmaster model. Its frozen AAC-style test top-five accuracy after two characters is 85.17%, compared with 78.11% for the earlier model. General English improves from 65.59% to 66.40%; Taskmaster falls from 86.56% to 86.24%. These are offline corpus measurements, not observed AAC-user outcomes.

Download and use

Choose a successful run for the desired commit in GitHub Actions. Download english-prediction-database and the cli-... artifact for your OS. Each contains a ZIP and archive checksum. The CLI ZIP filename identifies its architecture. Extract the CLI and model into separate directories, verify the checksums, and run each bundle's python verify_bundle.py.

/path/to/switchify-prediction validate --database /path/to/model/english.sqlite --production
/path/to/switchify-prediction predict --baseline /path/to/model/english.sqlite \
  --before 'I need ' --prefix 'he'

On Windows use switchify-prediction.exe. The executable includes SQLite; no server or Rust installation is needed. CLI packages are unsigned, not desktop app installers. Actions downloads require GitHub login and expire after 90 days. Approved version tags run the same validation and prepare a draft release for durable distribution; see release and operation guidance. No network access is needed for prediction or learning.

Prediction returns JSON with word and score. Defaults are five suggestions after two Unicode grapheme clusters. Override with --limit and --min-chars; use --min-chars 0 for next-word prediction. Pass preceding completed text in --before and the unfinished current word separately in --prefix. Suggestions are lowercase; insertion and capitalization are the application's responsibility. Scores are interpolated probabilities before prefix filtering, not renormalized over the returned suggestions or calibrated confidence estimates.

Build reproducibly

Requires Rust 1.97.1, a C compiler for bundled SQLite, Python 3.9+ and curl. Run from this checkout:

cargo build --release --locked
python3 scripts/artifacts.py

This fetches checksum-pinned public data, prepares the frozen partitions, builds the comparison models, evaluates development and test, checks the frozen quality results, then produces the production bundle in artifacts/model-bundle and unpacked files in artifacts/production. It never reads personal data. The OANC download is about 655 MB. Cached verified sources allow offline reruns.

To prepare and build without repeating the quality benchmark:

python3 scripts/aac_experiment.py --prepare-only
mkdir -p artifacts
target/release/switchify-prediction build --output artifacts/english.sqlite

The default build requires the exact prepared training text and verifies the model fingerprint before publishing. It embeds full source and policy provenance in SQLite. build --input training.txt --output custom.sqlite instead builds a custom model from explicit UTF-8 text. Existing output files are never overwritten. Plain validate checks any compatible model; validate --production additionally requires this binary's exact shipped model and provenance. A custom build does not inherit the official baseline's evaluation or distribution permissions.

Personal learning

switchify-prediction import --baseline english.sqlite \
  --personal personal.sqlite --input training.txt
printf 'I would like watermelon.' | switchify-prediction learn \
  --baseline english.sqlite --personal personal.sqlite
switchify-prediction predict --baseline english.sqlite --personal personal.sqlite \
  --before 'I would like ' --prefix 'wa'
switchify-prediction reset-personal --baseline english.sqlite --personal personal.sqlite

Imports deduplicate exact UTF-8 content by SHA-256. Every learn call represents one completed, non-overlapping segment; never repeatedly submit a growing typing buffer. Context does not cross calls, newlines, ., ! or ?. Internal apostrophes are preserved, curly apostrophes normalize to ASCII, and keys use Unicode NFC and lowercase. Other punctuation separates words. Numbers are ignored.

Missing personal databases are initialized atomically; invalid existing databases are rejected. Writes are transactional, including validation of the new in-memory state before commit. Multiple writers serialize through SQLite with a five-second busy timeout. Each predictor uses a snapshot; call refresh_personal() to see another instance's updates. Successful writes and duplicate imports refresh the calling instance. Errors preserve its previous snapshot.

Personal n-grams can reveal sensitive content and are stored unencrypted. Keep these files in an access-controlled application data directory, outside source control and distribution bundles. Reset removes logical counts and import markers, not forensic copies or backups. The library does not upload or log training text.

Rust API

use std::path::Path;
use switchify_prediction::{Options, Predictor, Result};

fn example() -> Result<()> {
    let mut predictor = Predictor::open(
        Path::new("english.sqlite"), Some(Path::new("personal.sqlite")))?;
    let suggestions = predictor.predict("I would like ", "wa", Options::default());
    predictor.learn("I would like watermelon.")?;
    predictor.refresh_personal()?;
    assert!(suggestions.len() <= 5);
    Ok(())
}

build, validate, production::build_english, production::validate_english, Predictor::import and Predictor::reset_personal cover the other operations. Fallible operations return typed Error values. Reuse a long-lived predictor; keep loading, imports, learning, refresh and resets on a worker thread. Prediction uses memory only. The model uses unigram/bigram/trigram weights 0.1/0.3/0.6, renormalized over available contexts, and combines baseline counts with 5× personal counts before probability calculation. Prefix matches rank by score, then alphabetically. Schema v1 and existing personal databases remain compatible.

Validation and model selection

cargo fmt --all -- --check
cargo clippy --locked --all-targets -- -D warnings
cargo test --locked
python3 -m unittest discover -s scripts -p 'test_*.py'
cargo audit --deny warnings

Install the CI-pinned auditing tool with cargo install cargo-audit --locked --version 0.22.0. CI runs these checks and packaged CLI/model smoke tests on macOS, Windows and Linux (dependency audit runs on Linux). Tests use temporary files and synthetic personal text.

production-model.json pins the accepted training and logical count hashes, parameters, source records and partition hashes. production-quality.json preserves the accepted quality measurements. Production artifacts require both AAC development/test gates to pass and every frozen accuracy/coverage/selection result to match. Warm latency is reported against a 20 ms p95 target, not used as a timing-sensitive CI gate. The accepted candidate measured below 6 ms p95 on an Apple M2 Max.

The training mixture is the prior 3× WorldAlphabets / 1× Taskmaster USER model, plus 10× each of 4,299 eligible unique AAC training sentences and 1× each of 20,000 hash-selected short OANC spoken utterances. Held-out text stays excluded. Official AAC worker and Taskmaster conversation splits are preserved; exact normalized sentence overlap is removed across training and evaluation.

The model bundle contains quality-report.json, validation, source/partition records, checksums and attribution. The evaluated candidate's file hash differs from the production file because production adds full provenance; their logical counts are identical and checked. The selection-saving metric is an optimistic rank-sensitive proxy, not measured switch savings. AAC is imagined by crowd workers, OANC is older American speech and Taskmaster is simulated task dialogue. Near-duplicates, domain bias, misspellings and inappropriate suggestions can remain. English alone is supported.

The AAC experiment protocol and earlier conversational results document model-selection history. The prepare and evaluate commands retain the earlier comparison workflows; prepare alone does not prepare the promoted production corpus. Use the commands above for production. A new model-selection exercise needs a new protocol; these test sets are now known regression sets.

New code is MIT licensed. Corpus licences are separate: retain ATTRIBUTION.md, source manifests and corpus-notices with model distributions. Production guidance covers the recorded licensing evidence, OANC's differing historical/current notices, private-data handling, backup, upgrades, rollback and releases.