Semantic Dev

Semantic Development

Zero-dependency Go library for matching natural language queries against accessibility tree elements.

Essential Commands

Before any PR:

./dev pr                # runs: check + e2e + lint corpus + bench

During development:

./dev test              # unit tests (fast)
./dev check             # fmt + vet + lint + test race (full validation)
./dev build             # build ./semantic CLI binary

Quality regression checks:

./dev baseline check    # compare quality against baseline
./dev runtime           # compare performance against baseline

When quality changes intentionally:

./dev baseline update   # accept new quality baseline (after review)

When to Use Each

Architecture

semantic.go               Public API (type aliases + constructors)
semantic_test.go           API-level smoke tests

internal/
  types/types.go           Type definitions (interfaces, structs)
  engine/                  All matching implementations (hidden)
    combined.go              Fused lexical + embedding matcher
    lexical.go               Jaccard similarity + synonyms + role boosting
    embedding.go             Cosine similarity on dense vectors
    hashing.go               Feature hashing embedder (zero-dep)
    synonyms.go              54 synonym groups for UI vocabulary
    stopwords.go             Context-aware stopword removal

recovery/                  Public subpackage
  engine.go                  RecoveryEngine (semantic re-matching)
  cache.go                   IntentCache (per-tab LRU)
  failure.go                 FailureType classification

cmd/semantic/main.go       CLI tool (find, match, classify)
cmd/semantic-bench/        Benchmark CLI (check, baseline, calibrate, tune, runtime)

Key Design Decisions

• internal/ — implementations are hidden. Consumers only see ElementMatcher interface + constructors.
• recovery/ — public subpackage. Pinchtab imports both semantic and semantic/recovery.
• Zero dependencies — hashing embedder, no ML models, no network calls.
• Stateless — every Find() call is independent. Thread-safe by default.

Workflow: New Feature or Bug Fix

1. Run doctor first:

   ./dev doctor
   

2. Make changes — implementations go in internal/engine/, types in internal/types/

3. Run checks:

   ./dev check   # fmt + vet + lint + test with race
   

4. Pre-commit hook runs gofmt + golangci-lint automatically on staged files.

Benchmark Improvement Loop

When implementing changes that affect matching quality:

./dev baseline          # create baseline (first time only)
# ... make changes ...
./dev bench             # run benchmark, compare to baseline
./dev baseline update   # accept new baseline (if improved)

Key metrics:

• MRR — Mean Reciprocal Rank (higher = finds correct element faster)
• P@1 — Precision at 1 (is top result correct?)
• Hit@3 — Any correct result in top 3?

Adding test cases:

1. Add to tests/benchmark/corpus/*/queries.json
2. Run ./dev lint corpus to validate
3. Run ./dev bench — shows regression until fixed

Public API Surface

Only these symbols are visible to consumers:

// Types (from internal/types via aliases)
semantic.ElementMatcher      // interface
semantic.Embedder            // interface
semantic.ElementDescriptor   // struct
semantic.ElementMatch        // struct
semantic.FindOptions         // struct
semantic.FindResult          // struct

// Constructors
semantic.NewCombinedMatcher(embedder)  → ElementMatcher
semantic.NewHashingEmbedder(dim)       → Embedder
semantic.NewLexicalMatcher()           → ElementMatcher
semantic.NewEmbeddingMatcher(e)        → ElementMatcher

// Functions
semantic.CalibrateConfidence(score)    → string
semantic.LexicalScore(query, desc)     → float64
semantic.CosineSimilarity(a, b)        → float64

// Recovery subpackage
recovery.IntentCache, recovery.IntentEntry
recovery.RecoveryEngine, recovery.RecoveryResult
recovery.ClassifyFailure, recovery.DefaultRecoveryConfig

Testing

• 167 tests across 3 packages (root, engine, recovery)
• internal/engine/ has unit tests for all matchers + benchmark suite
• Root has API-level smoke tests
• recovery/ has scenario tests (SPA re-render, checkout, login, etc.)

Release Process

1. Update npm/package.json version
2. git tag v0.X.0 && git push origin v0.X.0
3. CI builds binaries (GoReleaser) + opens homebrew tap PR
4. npm publish manually or via release-manual-publish.yml