convez.ai
SkillDados & AnalyticsAvançado
miketung

Bayes — Redes Bayesianas com Interface Web e Skill Claude

miketung/bayes

Redes bayesianas com web UI, CLI e integração nativa com Claude

Ver no GitHub
BomPluginScore · de 0 a 100

Toolkit completo para modelagem e inferência em redes bayesianas, com biblioteca JavaScript zero-dependências, interface web drag-and-drop para edição de DAGs, e skill CLI para integração com Claude Code. Permite construir modelos probabilísticos estruturados, fazer consultas de probabilidade marginal e usar o AI enrichment para preencher tabelas de probabilidade condicional via web search.

Instalação

$npx convez-ai miketung/bayes

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
6
Forks
1
Acompanhando
1
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Modelagem de raciocínio probabilístico estruturado em conversas Claude
  • Ferramentas educacionais para ensino de teoria de probabilidades e modelos gráficos
  • Diagnóstico e análise de cenários multi-variável com dependências complexas

Veredito

Projeto maduro e bem-arquitetado que une rigor matemático com UX acessível. A integração native com Claude Code e o suporte a AI enrichment a distinguem, mas exige familiaridade com probabilidades e DAGs; não é ferramenta casual.

README

A tradução para o português deste README ainda está na fila.

Bayes

The theory of probabilities is at bottom nothing but common sense reduced to calculus; it enables us to appreciate with exactness that which accurate minds feel with a sort of instinct for which ofttimes they are unable to account.

— Pierre Simon Laplace, 1819

A Bayesian network toolkit in three parts:

  • lib/ — a zero-dependency ES-module JavaScript library. Runs in the browser and in Node. Discrete networks, variable elimination, likelihood weighting, JSON I/O.
  • web/ — a static SPA: drag-and-drop DAG editor, CPT table editor, live marginals, mobile-friendly list view. Modeled after BayesFusion's GeNIE/BayesBox/BayesMobile.
  • cli/ — a Node CLI + Claude Code skill that lets Claude treat a Bayesian network as a structured reasoning/memory tool.

Everything shares one package.json and the same library code.

Quick start

npm install
npm run dev       # runs the server at http://localhost:3001 (with --watch)
npm test          # runs the inference / IO test suite

npm run dev starts a single Node server that serves the static SPA and the AI enrichment API. The static app works fine on its own — the AI button only appears when an API key is configured.

Optional AI enrichment

If you'd like the "✦ AI fill" button in the node inspector to search the web and fill in node probabilities and citations automatically, copy .env.example to .env and set OPENAI_API_KEY:

cp .env.example .env
echo "OPENAI_API_KEY=sk-..." >> .env
npm run dev

Override the base URL (OPENAI_API_BASE) to point at any OpenAI-compatible endpoint that supports the Responses API + web_search tool. Model is configurable via BAYES_MODEL (default: gpt-5.4-mini).

When you click ✦ AI fill in the inspector, the backend runs the web_search tool through the LLM, synthesizes a marginal (for root nodes) or a full CPT (for nodes with parents), and attaches the cited sources to the node. Sources are persisted with the network JSON so you can see them again next time you load the file.

Using the CLI

# interactive use from the repo
node cli/bayes.js query examples/asia.json --id dysp --format text

# or install globally
npm link
bayes list examples/earthquake.json --format text
bayes set-evidence examples/earthquake.json --id john --state yes --out /tmp/eq.json
bayes query /tmp/eq.json --id burglary --format text

Full command list: bayes help. Or see skill/SKILL.md for a worked example and a description of the JSON schema.

Installing the Claude Code skill

# make the CLI available on PATH
npm link

# install the skill so Claude Code picks it up
mkdir -p ~/.claude/skills/bayes
cp skill/SKILL.md ~/.claude/skills/bayes/SKILL.md

From then on, a Claude Code session can say "model this with a Bayes net" and Claude will shell out to bayes to build/query a network on disk.

Library API

import { BayesNet, parse, stringify, infer } from './lib/index.js';

const net = new BayesNet('demo');
net.addNode({ id: 'rain',  states: ['no','yes'] });
net.addNode({ id: 'wet',   states: ['no','yes'], parents: ['rain'] });
net.setCPT('rain', [0.8, 0.2]);
net.setCPT('wet',  [0.9, 0.1, 0.1, 0.9]);   // row-major, parents most-sig first
net.setEvidence('wet', 'yes');

infer(net, 'rain');                          // { no: 0.31, yes: 0.69 }
infer(net, 'rain', { algorithm: 'lw', samples: 20000 });

CPT layout

For a node with parents [P1, P2, …, Pk] (cardinalities [c1, …, ck]) and m own states, the CPT is a flat row-major array of length c1·c2·…·ck·m. Each row of length m is a distribution summing to 1. Rows iterate parent combinations with P1 as the most significant digit, Pk as the least.

Within a row, entries are in the declared order of the node's own states.

JSON schema

{
  "version": 1,
  "name": "Asia",
  "nodes": [
    { "id": "asia", "name": "Visit to Asia", "states": ["no","yes"],
      "parents": [], "cpt": [0.99, 0.01] }
  ],
  "evidence": { "asia": "yes" },
  "positions": { "asia": {"x": 60, "y": 40} }
}

Web app features

  • Cytoscape-powered DAG canvas with pan / zoom / drag
  • Per-node label shows live marginal probabilities and tiny bar charts
  • Right-side inspector: rename, edit states, edit CPT, set evidence
  • Auto-layout via dagre
  • Three built-in example networks (Asia, Cancer, Earthquake)
  • Load / save JSON; ships with no backend
  • Switch between Variable Elimination (exact) and Likelihood Weighting (approx)
  • Mobile: off-canvas drawer, bottom-sheet inspector, BayesMobile-style list view
  • Tailwind via CDN; no build step required to hack on the UI

Interactions

  • Connect nodes: hover over a node — a handle appears on its edge. Drag the handle onto another node to create a directed arc. The Parents section in the inspector has a dropdown as a keyboard-free alternative.
  • Delete: click a node or edge to select it, then press Delete (or Backspace).
  • Edit: click a node to open the inspector. Rename, add/remove states, edit the CPT table, or set evidence. Everything re-runs inference live.

Repository layout

index.html              static entry (served by server/index.js)
lib/                    core library (zero deps, browser + node)
web/                    SPA wiring + UI components
cli/                    Node CLI (#!/usr/bin/env node, zero deps)
skill/                  Claude Code skill definition
examples/               classic networks in JSON
test/                   node --test test suite

Scope / roadmap

Included: discrete chance nodes, exact inference (VE), approximate inference (LW), JSON I/O, full graphical editor, mobile layout.

Not yet: decision nodes / utility nodes (influence diagrams), dynamic Bayesian networks, continuous / equation nodes, noisy-MAX / noisy-adder, sensitivity analysis, XDSL / Netica / BIF import. All are reachable from this foundation without redesigning the data model.

Ver o original no GitHub

Bom para

Relacionados