pm4wasm

Process mining in Rust/WebAssembly — a WASM port of pm4py with LLM integration for natural language process discovery.

pm4wasm brings the core pm4py algorithms to the browser, enabling process mining entirely client-side with no server required. Plus, LLM-powered features for generating process models from natural language descriptions.

Install

npm install pm4wasm

Quick start

import { Powl } from "pm4wasm";

const powl = await Powl.init();

// Parse POWL model
const model = powl.parse("PO=(nodes={A, B, C}, order={A-->B, A-->C})");
model.validate();                       // throws if invalid
console.log(model.toString());          // canonical string
console.log([...model.activities()]);   // ["A", "B", "C"]

// Convert to Petri net
const petriNet = model.toPetriNet();
console.log(petriNet.net.transitions.length);

// Parse event log from CSV
const log = powl.parseCsv(
  "case_id,activity\n1,A\n1,B\n1,C\n2,A\n2,C\n"
);

// Check conformance
const fitness = powl.conformance(model, log);
console.log(fitness.percentage);         // 0.0 – 1.0
console.log(fitness.perfectly_fitting_traces);

// Filter by fitness threshold
const goodTraces = powl.filterByFitness(model, log, 0.8);

pm4py WASM Port

pm4wasm ports the core pm4py process mining algorithms to WebAssembly:

Discovery (from pm4py)

• discoverDFG() — Directly-Follows Graph
• discoverDFGTyped() — Typed DFG object format
• discoverPerformanceDFG() — Performance DFG with duration stats
• discoverEventuallyFollowsGraph() — Eventually-follows relations
• discoverProcessTree() — Inductive process tree
• discoverPetriNet() — Inductive Petri net
• discoverPetriNetAlpha() — Alpha miner
• discoverPetriNetAlphaPlus() — Alpha+ miner (handles loops)
• discoverPrefixTree() — Trie (prefix tree) of trace prefixes
• discoverLogFootprints() — Footprints discovery

Conformance (from pm4py)

• conformance() — Token-replay fitness
• conformancePetriNet() — Fitness on pre-built Petri net
• conformanceFootprints() — Footprints-based fitness/precision/recall/f1
• checkSoundness() — Deadlock freedom, liveness, boundedness

Filtering (from pm4py)

• 15 log filters: start/end activities, variants, time range, attributes, case size, prefixes/suffixes, and more

Statistics (from pm4py)

• 20+ statistics: start/end activities, variants, case durations, rework times, overlaps, performance stats, and more

I/O (from pm4py)

• XES and CSV read/write
• BPMN 2.0 XML export
• JSON-OCEL read/write (object-centric event logs)

OCEL Support

• parseOcelJson() — Parse OCEL 1.0/2.0 JSON
• ocelFlattenByObjectType() — Flatten to traditional log by object type
• discoverOcelEtot() — Event-Type / Object-Type graph
• ocelGetSummary() — OCEL statistics
• ocelGetObjectTypes() — List all object types
• ocelGetEventTypes() — List all event types

LLM Integration (Beyond pm4py)

Generate POWL models from natural language using Groq, OpenAI, or Anthropic:

// Natural language → POWL
const model = await powl.fromNaturalLanguage(
  "Customer orders and pays, then receives confirmation",
  {
    provider: "groq",
    apiKey: process.env.GROQ_API_KEY,
  },
  "ecommerce"
);

// POWL → BPMN
const bpmn = powl.toBpmn(model.toString());

// Generate workflow code directly
const n8nWorkflow = await powl.naturalLanguageToCode(
  "Order processing with payment",
  "n8n",
  { provider: "groq", apiKey: process.env.GROQ_API_KEY }
);

Supported LLM Providers

| Provider | Speed | Cost | Best For | |----------|-------|------|----------| | Groq | ⚡⚡⚡ | Free | Development, fast iteration | | OpenAI | ⚡⚡ | Paid | Production, GPT-4o | | Anthropic | ⚡ | Paid | Claude 3.5 Sonnet |

Build from source

Requires Rust + wasm-pack.

# Install wasm-pack (once)
curl https://rustwasm.github.io/wasm-pack/installer/init.sh -sSf | sh

# Build WASM then TypeScript
cd js/
npm install
npm run build

Dev / demo server

npm run demo
# Opens http://localhost:5173 — live POWL editor + conformance checker

Run browser tests

npm test                   # Firefox headless
# or:
wasm-pack test .. --headless --chrome

API overview

Powl.init() → Promise<Powl>

Loads the WASM module once; safe to call multiple times.

Parsing

| Method | Description | |--------|-------------| | powl.parse(str) | Parse a POWL model string | | powl.parseXes(xml) | Parse a XES event log | | powl.parseCsv(csv) | Parse a CSV event log | | powl.fetchXes(url) | Fetch + parse XES from URL | | powl.readXesFile(file) | Parse File drag-drop XES | | powl.readCsvFile(file) | Parse File drag-drop CSV |

PowlModel

| Method | Description | |--------|-------------| | .toString() | Canonical model string | | .validate() | Throws on SPO violations | | .simplify() | Structure-normalized model | | .simplifyFrequent() | Convert XOR/LOOP+tau → FrequentTransition | | .toPetriNet() | Returns PetriNetResult | | .nodeInfo(idx) | Typed node description | | .children(idx) | Child arena indices | | .activities() | All activity labels | | .walk(visitor) | Pre-order tree traversal | | .orderEdges(idx) | SPO ordering relation edge list | | .closureEdges(idx) | Transitive closure edge list | | .reductionEdges(idx) | Transitive reduction edge list |

Conformance

| Method | Description | |--------|-------------| | powl.conformance(model, log) | Token-replay fitness | | powl.conformancePetriNet(pn, log) | Fitness against pre-built Petri net | | powl.filterByFitness(model, log, threshold) | Filter traces by fitness | | powl.variants(log) | Variant frequency map |