Skip to content

002-intent-recognition.md

Latest commit

 

History

History
79 lines (62 loc) · 2.14 KB

002-intent-recognition.md

File metadata and controls

79 lines (62 loc) · 2.14 KB
Raw

ADR 002: Intent Recognition Strategy

Status

Accepted

Context

We needed an intent recognition system that:

  • Works offline without external APIs
  • Provides confidence scores for matching
  • Handles variations in user phrasing
  • Is fast enough for real-time conversation
  • Can be tuned per scenario

Decision

We implemented a multi-metric similarity matching approach combining:

1. TF-IDF + Cosine Similarity

// Build IDF cache from all intent examples
const tfidf = computeTFIDF(tokens);
const similarity = cosineSimilarity(inputTFIDF, exampleTFIDF);

2. Jaccard Similarity

const jaccard = jaccardSimilarity(inputTokens, exampleTokens);

3. Keyword Boost

// Boost score for critical keywords
const keywordBoost = calculateKeywordBoost(input, pattern);

4. Weighted Combination

const finalScore = (cosine * 0.5) + (jaccard * 0.3) + (keywordBoost * 0.2);

5. Fallback Chain

  1. Exact match (button clicks, exact phrases)
  2. Intent-based matching (multi-metric)
  3. Regex patterns
  4. Keyword matching
  5. Fuzzy string matching (Levenshtein)

Consequences

Positive

  • ✅ 89% intent accuracy on test scenarios
  • ✅ ~5ms average matching time
  • ✅ Works completely offline
  • ✅ Confidence thresholds allow graceful degradation
  • ✅ Emotion detection adds contextual awareness

Negative

  • ❌ Requires example phrases for each intent
  • ❌ May miss context that LLMs would catch
  • ❌ Performance degrades with very large intent sets

Mitigations

  • Confidence thresholds prevent low-quality matches
  • Fallback to fuzzy matching catches edge cases
  • Intent sets are scenario-specific (typically 5-15 intents)

Alternatives Considered

Option A: External LLM API

  • Rejected due to cost, latency, and offline requirement

Option B: Pure Regex Matching

  • Rejected due to inflexibility and maintenance burden

Option C: Naive Bayes Classifier

  • Considered but TF-IDF provides better accuracy for short text

References