kjarni

Local AI inference for C#, Go and the command line

Classification, embeddings, semantic search, and reranking. One native library. No Python. No CUDA. No containers. Reads from stdin, writes JSON, pipes like any UNIX tool.

Hybrid search with reranking, semantic + BM25

Text classification, no API keys, no cloud, runs locally

Embeddings and similarity

dotnet add package Kjarni

go get github.com/olafurjohannsson/kjarni-go@latest

curl -fsSL https://kjarni.ai/install.sh | sh

irm https://kjarni.ai/install.ps1 | iex

Zero dependencies

Models auto-download

Works offline

Reads from stdin

Written in Rust

What It Does

Four capabilities. Same API in C# and on the command line.

Classification

Sentiment, emotion, toxicity detection. Pass text in, get labels with confidence scores.

DistilBERT · RoBERTa · BERT Multilingual · Toxic-BERT

Embeddings

Dense vector representations for semantic similarity, clustering, and retrieval.

MiniLM-L6 · MPNet · DistilBERT

Semantic Search

Index files from a directory. Keyword, semantic, and hybrid retrieval built in.

HNSW · BM25 · Hybrid

Reranking

Cross-encoder scoring to reorder search results and improve retrieval quality.

MiniLM Cross-Encoder

Why Kjarni?

You shouldn't need a PhD to classify an email.

Self-Contained

One install, nothing else. No Python runtime, no CUDA toolkit, no Docker containers, no API keys. Models download on first use and cache locally.

dotnet add package Kjarni
cargo install kjarni-cli
Works offline after first run

Task-Level API

You get Classifier, not BertForSequenceClassification. Kjarni hides tokenizers, attention masks, and pooling strategies.

Classifier, Embedder, Searcher, Reranker
6 classifiers, 3 embedders, cross-encoder
Pick a model name, call a method

UNIX-Native

The CLI reads from stdin, writes to stdout, and outputs JSON. Pipe it to jq, grep, or into your scripts. It's a tool, not a framework.

cat reviews.txt | kjarni classify
kjarni classify --format json | jq
SIMD-optimized (AVX2, NEON)

Three Lines of Code

Same capabilities in C#, Go, or the terminal.

C# Sentiment Analysis

using Kjarni;

var clf = new Classifier("distilbert-sentiment");
var result = clf.Classify("Best purchase I've ever made!");

Console.WriteLine($"{result.Label}: {result.Score:P}");
// POSITIVE: 100.0%

Go Sentiment Analysis

import "github.com/olafurjohannsson/kjarni-go"

c, _ := kjarni.NewClassifier("distilbert-sentiment")
defer c.Close()

result, _ := c.Classify("Best purchase I've ever made!")
fmt.Printf("%s: %.1f%%\n", result.Label, result.Score*100)
// POSITIVE: 100.0%

CLI Sentiment Analysis

$ kjarni classify "Best purchase I've ever made!"
  ✓       POSITIVE  ████████████████████  100.0%
          NEGATIVE  ░░░░░░░░░░░░░░░░░░░░    0.0%

# Pipe from stdin
$ echo "Terrible quality" | kjarni classify
  ✓       NEGATIVE  ████████████████████  100.0%
          POSITIVE  ░░░░░░░░░░░░░░░░░░░░    0.0%

C# Toxicity Detection

var clf = new Classifier("toxic-bert");
var result = clf.Classify(userMessage);

if (result.Label == "toxic" && result.Score > 0.8f)
{
    // flag for moderation
}
// Multi-label: toxic, insult, obscene, threat, ...

Go Toxicity Detection

c, _ := kjarni.NewClassifier("toxic-bert")
defer c.Close()

result, _ := c.Classify(userMessage)

if result.Label == "toxic" && result.Score > 0.8 {
    // flag for moderation
}
// Multi-label: toxic, insult, obscene, threat, ...

CLI Toxicity Detection

$ kjarni classify "You are the worst cook ever" --model toxic-bert
  ✓          toxic  ███████████████████░   92.8%
            insult  ██████████████░░░░░░   72.3%
           obscene  ██░░░░░░░░░░░░░░░░░░    7.6%
     identity_hate  ░░░░░░░░░░░░░░░░░░░░    0.5%
      severe_toxic  ░░░░░░░░░░░░░░░░░░░░    0.3%

C# Embeddings & Similarity

var emb = new Embedder("minilm-l6-v2");

float sim = emb.Similarity("doctor", "physician");
Console.WriteLine($"Similarity: {sim:P}");
// Similarity: 86.0%

Go Embeddings & Similarity

e, _ := kjarni.NewEmbedder("minilm-l6-v2")
defer e.Close()

sim, _ := e.Similarity("doctor", "physician")
fmt.Printf("Similarity: %.1f%%\n", sim*100)
// Similarity: 86.0%

CLI Embeddings & Similarity

$ kjarni similarity "doctor" "physician"
  █████████████████░░░   86.0%  highly similar

$ kjarni similarity "doctor" "banana"
  ███████░░░░░░░░░░░░░   33.8%  somewhat related

C# Document Search

var indexer = new Indexer("minilm-l6-v2");
indexer.Create("./my-index", new[] { "./docs" });

var searcher = new Searcher("minilm-l6-v2", "");
var results = searcher.Search("./my-index", "query");

foreach (var r in results)
    Console.WriteLine($"{r.Score:F3}: {r.Text}");

Go Document Search

idx, _ := kjarni.NewIndexer("minilm-l6-v2")
defer idx.Close()
idx.Create("./my-index", []string{"./docs"})

s, _ := kjarni.NewSearcher("minilm-l6-v2", "")
defer s.Close()
results, _ := s.Search("./my-index", "query", kjarni.Hybrid)

for _, r := range results {
    fmt.Printf("%.3f: %s\n", r.Score, r.Text)
}

CLI Index & Search

# Index a folder of documents
$ kjarni index create my-docs docs/*
✓ Indexed 15 documents (39.52 KB)

# Search with hybrid retrieval
$ kjarni search my-docs "keeping data safe" --top-k 3
  1. cryptocraphy.txt
     ████████████████████  100.0%
  2. tcpip.txt
     ██████████░░░░░░░░░░   49.2%
  3. neuralnetworks.txt
     ░░░░░░░░░░░░░░░░░░░░    0.0%