What It Does
Extract
Feed it 5+ text samples → get a structured persona (7 voice dimensions)
Compile
Turn a persona into a compact prompt block for any LLM
Filter
Post-generation voice filter catches AI-sounding output (60+ patterns)
Track
Voice DNA tracks distinctive phrases confirmed by reuse
The 7 Voice Dimensions
| Dimension | Example |
|---|
| Tone | sharp, witty, blunt |
| Humor | dry sarcasm, never forced |
| Vocabulary | technical, casual |
| Formality | informal, no hedging |
| Perspective | pragmatic minimalist |
| Emotional range | controlled intensity |
| Sentence length | short, punchy |
- Minimum 3 samples required, 5+ recommended
- Diversity scoring via Type-Token Ratio — detects if samples are too similar
- Confidence score (0.0–1.0) based on sample count (30%), diversity (40%), and content richness (30%)
- Warnings for low-quality extractions
use laminae::persona::{PersonaExtractor, VoiceFilter, VoiceFilterConfig, compile_persona};
let extractor = PersonaExtractor::new("qwen2.5:7b");
let persona = extractor.extract(&samples).await?;
// Check extraction quality
if let Some(quality) = &persona.meta.quality {
println!("Confidence: {:.0}%", quality.confidence * 100.0);
println!("Diversity: {:.0}%", quality.diversity_score * 100.0);
for warning in &quality.warnings {
eprintln!("Warning: {warning}");
}
}
let prompt_block = compile_persona(&persona);
Voice Filter
The voice filter catches AI-sounding output with 6 detection layers:
| Layer | What It Catches |
|---|
| AI vocabulary | 60+ phrases like “it’s important to note”, “furthermore”, “delve” |
| Meta-commentary | ”The post highlights…” openers |
| Multi-paragraph | Optional rejection of multi-paragraph responses |
| Trailing questions | Generic AI questions at the end |
| Em-dashes | Replace with periods |
| Length violations | Sentence and character limits |
let filter = VoiceFilter::new(VoiceFilterConfig::default());
let result = filter.check("It's important to note that...");
// result.passed = false
// result.violations = ["AI vocabulary detected: ..."]
// result.retry_hints = ["DO NOT use formal/academic language..."]
Voice DNA
Tracks distinctive phrases that define a person’s writing. When generated text receives positive feedback, distinctive phrases are recorded and reinforced.
use laminae::persona::VoiceDna;
let mut dna = VoiceDna::empty();
dna.record_success("Nobody reads your README. Ship it or delete it.");
// "Nobody reads" is now tracked as a distinctive phrase
The voice filter is WASM-compatible and available via Python bindings. The extractor requires Ollama (native only).
