kiwi-rs-assistant

name	kiwi-rs-assistant
description	Guide AI to use kiwi-rs accurately for Korean NLP in Rust. Use when tasks involve kiwi-rs setup, init/new/from_config selection, analyze/tokenize/split/join APIs, builder dictionary or typo customization, UTF-16 or batch APIs, kiwipiepy migration, or kiwi-rs runtime troubleshooting.

Kiwi Rs Assistant

Use this skill to produce runnable, minimal, and correct kiwi-rs solutions.

Workflow

1. Identify intent first.

• Map the request to one primary task: tokenize, analyze, split, join/space/glue, builder customization, typo correction, UTF-16 path, batch processing, or semantics APIs.
• Prefer the smallest API surface that satisfies the request.

2. Choose initialization path deliberately.

• Use Kiwi::init() for the easiest path with automatic bootstrap.
• Use Kiwi::new() when KIWI_LIBRARY_PATH and KIWI_MODEL_PATH are already managed by environment.
• Use Kiwi::from_config(KiwiConfig::default()...) when explicit, reproducible paths are required.
• Use KiwiLibrary + builder(...) when adding user words, regex rules, dictionaries, or typo sets before building Kiwi.

3. Generate runnable Rust code.

• Return complete snippets that compile (fn main() -> Result<(), Box<dyn std::error::Error>>).
• Import only required items from kiwi_rs.
• Keep text examples Korean unless the user requests otherwise.
• If the user asks for many features, stage the answer as a minimal baseline plus one extension block.

4. Add one validation step.

• Provide at least one command that confirms behavior (cargo run --example ... or cargo run for the snippet project).
• Include expected high-level output shape (for example: token count, candidate count, sentence boundaries).

5. Add focused troubleshooting guidance.

• Map failure messages to concrete fixes using references/troubleshooting.md.
• If bootstrap is used, mention required external commands: curl, tar (and powershell on Windows).

Accuracy Guardrails

• Treat UTF-8 offsets as character indices (str.chars()), not byte indices.
• Check supports_utf16_api() before UTF-16 tokenization/sentence APIs.
• Check supports_analyze_mw() before analyze_many_utf16_via_native.
• Do not claim full kiwipiepy parity. Call out missing or partial areas (for example Template, Stopwords, and other Python/C++-specific layers).
• For deterministic spans or constraints, prefer MorphemeSet and Pretokenized.
• When discussing custom vocabulary or typo behavior, configure it in builder phase and rebuild Kiwi.

Response Contract

Always structure answers in this order:

1. Path: chosen initialization and why.
2. Code: copy-paste runnable snippet.
3. Verify: one concrete command.
4. Pitfalls: only the risks relevant to the user request.

References

• Read references/task-to-api.md to map user goals to APIs and local examples.
• Read references/troubleshooting.md when errors, environment issues, or parity confusion appear.