AI-Native Features

Built-in capabilities for production AI applications: detect hallucinations, verify facts, score confidence, and optimize prompts automatically.

Prompt Optimization

Automatically optimize prompts for better results and lower costs

import { optimizePrompt, compressPrompt } from '@rana/core';

// Compress a verbose prompt
const compressed = compressPrompt(longPrompt);
// Result: 40% fewer tokens, same meaning

// Optimize for a specific goal
const optimized = await optimizePrompt(prompt, {
  goal: 'quality',
  strategy: 'chain-of-thought'
});

Hallucination Detection

Detect fabricated facts, fake citations, and logical inconsistencies

import { detectHallucinations } from '@rana/core';

const result = detectHallucinations(response, {
  knownFacts: [...],
  context: sourceDocument
});

if (result.hasHallucinations) {
  console.log('Issues found:', result.instances);
  // Types: fabricated_citation, overconfidence,
  //        temporal_error, logical_inconsistency
}

Confidence Scoring

Measure response confidence through linguistic and consistency analysis

import { scoreConfidence, isConfident } from '@rana/core';

const score = scoreConfidence(response, {
  context: originalQuery,
  samples: [response1, response2, response3] // for consistency
});

console.log(score.overall);      // 0.85
console.log(score.level);        // 'high'
console.log(score.breakdown);    // { linguistic, consistency, specificity, grounding }
console.log(score.recommendations);

Fact Verification

Extract claims and verify them against knowledge bases

import { verifyFacts, extractClaims } from '@rana/core';

// Extract factual claims from text
const claims = await extractClaims(response);
// [{ text: "Paris is the capital of France", type: "factual" }]

// Verify claims
const result = await verifyFacts(response);
console.log(result.verifiedClaims);   // Claims with evidence
console.log(result.falseClaims);      // Number of false claims
console.log(result.overallReliability); // 0-1 score

Quality Scoring

Multi-dimensional quality evaluation for LLM responses

import { scoreQuality, getQualityLevel } from '@rana/core';

const quality = scoreQuality(response, query);

console.log(quality.overall);    // 0.87
console.log(quality.dimensions); // { relevance, completeness, clarity,
                                 //   accuracy, helpfulness, conciseness }
console.log(quality.suggestions); // Improvement recommendations

Comprehensive Analysis

Run all checks at once for complete response validation

import { analyzeResponse, isTrustworthy } from '@rana/core';

// Quick check
if (isTrustworthy(response)) {
  // Safe to use
}

// Full analysis
const analysis = await analyzeResponse(response, {
  query: originalQuery,
  context: { knownFacts: [...] }
});

console.log(analysis.overallScore);     // 0.82
console.log(analysis.hallucinations);   // Hallucination results
console.log(analysis.confidence);       // Confidence results
console.log(analysis.verification);     // Fact verification
console.log(analysis.quality);          // Quality scores
console.log(analysis.recommendations);  // Combined recommendations

Best Practices

Always verify critical information before presenting to users
Use confidence scoring to add uncertainty indicators in UI
Compress prompts in production to reduce costs by 30-50%
Log quality scores for monitoring and improvement tracking
Combine hallucination detection with fact verification for critical apps