The most comprehensive Ethiopic NLP toolkit
Fully composable, local-first rule stemmers, lexical normalizers, tokenizers, and loss-free ASCII transliteration. Powered by a high-performance Rust core with WASM and native Python bindings, running up to 35% faster than pure JavaScript implementations.
npm install @fidel-tools/core @fidel-tools/lang-am
# Initialize local pipeline
# import { Pipeline } from '@fidel-tools/core';
# import amPack from '@fidel-tools/lang-am';Interactive Execution Console
Type custom words or sentences to watch the pipeline execute normalizers, filters, and morphological rules.
import { Pipeline } from '@fidel-tools/core';
import amPack from '@fidel-tools/lang-am';
const nlp = new Pipeline(amPack);
const corpus = "የገንዘብ ሚኒስቴር ምክር ቤተ ከሃያ ዓመታት በፊ...";
const lexed = nlp.lexAnalyze(corpus);
const clean = nlp.removeStopwords(lexed);
const stems = clean.split(' ').map(w => nlp.stem(w));Platform Features
Designed with precision for natural language processing of Ethiopic typography, character sets, and stemming.
Lexical Normalization
Equates visual variations of character glyph configurations (e.g. ሃ/ሀ/ሐ, ኀ/ሀ, ሠ/ሰ, ዐ/አ) to maximize document indexing consistency.
Rule-Based Stemmer
Employs morphological patterns to strip inflected suffixes, prefixes, and infixes, yielding correct word roots.
Lossless Transliteration
Robust phonetic transliteration between standardized ASCII SERA phonetic strings and native Ge'ez scripts.
Tokenization Exception rules
Correctly handles compound abbreviations, custom delimiters, sentence end punctuation, and numbers.
Developer Console APIs
Manage credentials, monitor incoming usage streams, logs, metrics, and rate limit thresholds.
Multi-language bindings
Integrates with Hono, Next.js, Python SDK, or agentic frameworks via MCP configurations.
Closing the Ethiopic NLP Gap
Traditional NLP libraries are English-centric and fail when applied to Ge'ez-based scripts. Fidel Tools introduces a script-first, modular approach designed for maximum accuracy and zero overhead.
Power accurate search engine indexing (matching ልጆቻቸውን to ልጅ), reduce LLM token usage in Generative AI / RAG pipelines by stripping Amharic stopwords, and run translations offline.
Schema-First (JSON vs. Code)
In legacy systems, supporting a new regional language requires writing complex Python parsing classes. In Fidel Tools, it's just a JSON configuration file. Linguists define language specifications without coding, opening up rapid support for Tigrinya, Oromo, and Ge'ez.
Normalization-First Pipeline
Orthographic ambiguity is the silent killer of Ethiopic text analysis. Fidel Tools collapses spelling variations (e.g. ሀ/ሐ/ኀ) before tokenization, ensuring downstream search and ML engines read the exact same semantic intent.
Lightweight & Edge-Ready
No heavy neural networks or 500MB weights. The core rules engine is under 250KB. Run processing locally inside browser packages, serverless functions, or edge runtimes with zero API network latency.
Shared Script Adapters
Instead of reinventing the wheel, the core engine abstracts script mechanics (syllabary, unicode mappings) into a ScriptAdapter. Multiple regional languages share one adapter, allowing near-instant codebase extensibility.
Next-Gen Processing Speed
Benchmarks measured on standard medium-length paragraphs (~200 characters) comparing JS execution, WebAssembly, and native Python C-extensions.