@reaatech/classifier-evals-exporters

Status: Pre-1.0 — APIs may change in minor versions. Pin to a specific version in production.

Exporters for classifier evaluation results. Supports JSON (machine-readable), HTML (interactive report), Arize Phoenix (traces and embeddings), and Langfuse (observability traces).

Installation

npm install @reaatech/classifier-evals-exporters
# or
pnpm add @reaatech/classifier-evals-exporters

Feature Overview

• JSON export — machine-readable output with optional sample inclusion, PII-aware redaction
• HTML report — interactive SVG-based confusion matrix heatmap, per-class bar charts, metrics dashboard
• Phoenix export — Arize Phoenix trace export with full metrics as span attributes and HTTP transport
• Langfuse export — Langfuse trace ingestion with authentication, session grouping, and structured metadata
• Dual ESM/CJS output — works with import and require

Quick Start

import {
  exportToJson,
  exportToHtml,
  exportToPhoenix,
  exportToLangfuse,
} from "@reaatech/classifier-evals-exporters";

// JSON export (PII-redacted by default)
const jsonResult = exportToJson({ evalRun });
console.log(jsonResult.json);

// HTML report with confusion matrix and per-class metrics
const htmlResult = exportToHtml(evalRun, {
  includeConfusionMatrix: true,
  includePerClassMetrics: true,
  title: "Classifier v2 Evaluation",
});
console.log(htmlResult.html);

// Phoenix trace export
await exportToPhoenix({
  evalRun,
  options: { endpoint: "http://localhost:6006", datasetName: "intent-classifier-v2" },
});

// Langfuse trace export
await exportToLangfuse({
  evalRun,
  options: {
    publicKey: process.env.LANGFUSE_PUBLIC_KEY,
    secretKey: process.env.LANGFUSE_SECRET_KEY,
    traceName: "classifier-evaluation",
  },
});

API Reference

JSON Exporter

exportToJson(input: JsonExportInput): ExportResult

Exports an EvalRun as a structured JSON payload with configurable options.

const result = exportToJson({
  evalRun,
  options: { includeSamples: false },
});
// result.json → string, result.success → boolean

| Option | Type | Default | Description | |--------|------|---------|-------------| | includeSamples | boolean | false | Include raw sample data (PII consideration) | | includePerClass | boolean | true | Include per-class breakdown | | includeVisualizationData | boolean | false | Include visualization-ready data |

The JSON payload includes: run_id, dataset_name, dataset_path, total_samples, duration_ms, started_at, completed_at, metrics, confusion_matrix, gate_results (if present), judge summary (if judged), and redacted metadata.

HTML Exporter

exportToHtml(evalRun: EvalRun, options?: HtmlExportOptions): HtmlExportResult

Generates a self-contained interactive HTML report with inline SVG charts.

const report = exportToHtml(evalRun, {
  includeConfusionMatrix: true,   // SVG heatmap (default: true)
  includePerClassMetrics: true,   // Per-class metrics table (default: true)
  includeBaselineComparison: true, // Baseline comparison section (default: false)
  includeJudgeAnalysis: true,     // LLM judge results section (default: true)
  title: "My Evaluation Report",
});

The report includes:

• Header — dataset name, run ID, total samples, duration, timestamps
• Metrics grid — accuracy, macro/micro F1, precision, recall, MCC, Cohen's Kappa
• Confusion matrix — SVG heatmap with color-coded cells and labels
• Per-class metrics — table with precision, recall, F1, and support per label
• Gate results — pass/fail status for each gate (if gate results present)
• Judge results — agreement rate, cost breakdown (if judged)

SVG charts are generated inline — no external dependencies or CDN requests.

Phoenix Exporter

exportToPhoenix(input: PhoenixExportInput): Promise<ExportResult>

Publishes evaluation results to an Arize Phoenix server as OpenTelemetry traces.

await exportToPhoenix({
  evalRun,
  options: {
    endpoint: "http://localhost:6006",  // Phoenix endpoint (default)
    datasetName: "intent-classifier",   // Dataset name in Phoenix (default: "classifier-evals")
    apiKey: process.env.PHOENIX_API_KEY, // Optional API key
    metadata: { model: "v2", env: "staging" },
  },
});

The export creates a single trace per eval run with one span containing:

• Dataset info — name, path, total samples
• Full metrics — all 14 ClassificationMetrics as span attributes
• Confusion matrix metadata — labels array and class count
• Custom metadata — user-provided metadata merged with PII-redacted content

Uses fetch() with a 30-second timeout and configurable authentication via Authorization: Bearer header.

Langfuse Exporter

exportToLangfuse(input: LangfuseExportInput): Promise<ExportResult>

Publishes evaluation results to Langfuse as trace events.

await exportToLangfuse({
  evalRun,
  options: {
    publicKey: process.env.LANGFUSE_PUBLIC_KEY,
    secretKey: process.env.LANGFUSE_SECRET_KEY,
    baseUrl: "https://cloud.langfuse.com",  // Default
    traceName: "classifier-evaluation",       // Default
    sessionId: `eval-${Date.now()}`,
  },
});

The export creates a trace event with:

• Input — dataset path, total samples
• Output — full metrics, gate pass status
• Metadata — duration, run ID, confusion matrix class count

Uses HTTP Basic Authentication with publicKey:secretKey over fetch() with a 30-second timeout.

Usage Patterns

Export Pipeline

import { createEvalRunFromSamples } from "@reaatech/classifier-evals-metrics";
import { exportToJson, exportToHtml } from "@reaatech/classifier-evals-exporters";

// Build the eval run
const evalRun = createEvalRunFromSamples({ datasetPath: "./test.csv", samples });

// Export to JSON for CI artifacts
const jsonResult = exportToJson({ evalRun });
await fs.writeFile("results.json", jsonResult.json);

// Export to HTML for human review
const htmlResult = exportToHtml(evalRun, { title: "Production Eval — v2.1.0" });
await fs.writeFile("report.html", htmlResult.html);

PII-Safe Export

// JSON export redacts PII by default (includesSamples: false)
const safe = exportToJson({ evalRun });

// Metadata is automatically redacted for PII before inclusion
// Phoenix export also redacts metadata via redactObjectPII()

Related Packages

• @reaatech/classifier-evals — Core types and PII redaction
• @reaatech/classifier-evals-metrics — Classification metrics and eval run construction

License

MIT