Shell Autocomplete Demo

Real-time shell command autocomplete powered by a trained N-gram Markov model.

This is the Presentar showcase demo - demonstrating Zero-Infrastructure AI deployment with WASM.

Overview

The Shell Autocomplete demo loads a trained .apr model file and provides intelligent command suggestions as you type. No server required - everything runs in the browser via WebAssembly.

User Input → WASM Runtime → N-gram Model → Suggestions
     ↓
  "git c" → ["git commit", "git checkout", "git clone", ...]

Key Features

• Zero Infrastructure: No Python, no server, no cloud - pure WASM
• Real Trained Model: Uses aprender-shell-base.apr (not random weights)
• Sub-millisecond Inference: <1ms suggestion latency
• Dynamic Model Loading: Fetch models at runtime via fromBytes()
• 574KB Total Size: WASM binary with embedded model

Running the Demo

cd /home/noah/src/presentar
make serve
# Open http://localhost:8080/shell-autocomplete.html

Architecture

Model Format (APR)

┌──────────────────────────────────────────────────────────────┐
│ 32-byte Header                                               │
├──────────────────────────────────────────────────────────────┤
│ Magic: "APRN" (4 bytes)                                      │
│ Version: 1.0 (2 bytes)                                       │
│ Model Type: 0x0010 (N-gram LM)                               │
│ Metadata Size, Payload Size, Compression Type                │
├──────────────────────────────────────────────────────────────┤
│ Payload (zstd compressed, bincode serialized)                │
│ - N-gram counts: HashMap<context, HashMap<token, count>>     │
│ - Command frequencies: HashMap<command, frequency>           │
│ - Total command count                                        │
└──────────────────────────────────────────────────────────────┘

YAML Configuration

presentar: "1.0"
name: "shell-autocomplete"
version: "1.0.0"

models:
  shell:
    source: "./assets/aprender-shell-base.apr"
    format: "apr"

layout:
  type: "app"
  sections:
    - id: "input-section"
      widgets:
        - type: "autocomplete"
          id: "shell-input"
          placeholder: "Type a command..."
          model: "{{ models.shell }}"
          suggestions: "{{ models.shell | suggest(state.input, 8) }}"

Expression Language

The suggest transform enables model inference in expressions:

{{ models.shell | suggest(prefix, count) }}

Returns an array of suggestion objects:

{
  "suggestions": [
    {"text": "git commit", "score": 0.101},
    {"text": "git checkout", "score": 0.056}
  ]
}

Model Statistics

WASM API

Dynamic Loading (Recommended)

import init, { ShellAutocompleteDemo } from './pkg/presentar.js';

await init();

// Fetch model dynamically
const response = await fetch('./assets/aprender-shell-base.apr');
const bytes = new Uint8Array(await response.arrayBuffer());

// Create autocomplete with fetched model
const autocomplete = ShellAutocompleteDemo.fromBytes(bytes);

// Get suggestions
const result = JSON.parse(autocomplete.suggest("git ", 5));
console.log(result.suggestions);

Embedded Model (Testing)

// Uses model compiled into WASM binary
const autocomplete = new ShellAutocompleteDemo();

Files

Academic References

The N-gram model implementation is based on:

1. Chen & Goodman (1999). "An Empirical Study of Smoothing Techniques for Language Modeling"
2. Stolcke (2002). "SRILM - An Extensible Language Modeling Toolkit"

See the specification for complete references.

10X Competitive Advantage