v1.0.4

Bug Fix

Fix: WordExtractor prompt/schema contract inconsistency (#2)

The Dutch extraction prompt instructed the LLM to extract phraseological constructions as single units, but PartOfSpeech had no phrase value. When the LLM returned word_type="phrase", Pydantic validation failed and the entire extraction raised APIError — breaking save-to-dictionary flows in production.

Changes

• Added PHRASE = "phrase" to PartOfSpeech enum in core.models
• Added phrase to the allowed types list in the NL extraction prompt
• Added a few-shot example demonstrating phraseological unit extraction
• Regenerated nl.json prompt asset
• Added prompt-schema consistency tests to prevent future drift

v1.0.2

Highlights

• migrate repository and package docs to the canonical module-spec format
• add shared storage ports in core (ScoredPairProvider, RemoteProgressSyncPort) and the WordPairSnapshot sync DTO
• decouple database_cache from a hard-coded remote progress store while keeping ExerciseProgressStore as the default adapter
• align sampling, database, and cache docs and contracts around the new protocol-based storage boundary

v1.0.1

Major release for the multi-package monorepo migration.

Highlights:

• repository reorganized into package-local projects under packages/*
• each package now has its own pyproject.toml, ruff.toml, pytest.ini, tests/, and Makefile
• root make check now delegates into package-local make check
• shared DTOs such as WordPair and ScoredWordPair now live only in nl_processing.core.models
• legacy re-exports for shared models were removed
• aggregate root build was updated to package the multi-root layout correctly

Breaking changes:

• shared model imports must use nl_processing.core.models
• old single-tree repo layout and root test layout are no longer the source of truth

v0.5.0

What's New in 0.5.0

New Features

• WordSampler now supports swappable scored data sources via scored_store, enabling the same sampling logic to run against either remote ExerciseProgressStore or local DatabaseCacheService
• Added ScoredPairProvider Protocol to formalize the minimal sampler dependency: get_word_pairs_with_scores()
• Preserved backward compatibility: existing WordSampler(...) usage still constructs ExerciseProgressStore internally when scored_store is not provided

Documentation

• Updated sampling PRD and architecture docs to define Protocol-based injection and cache-backed sampling
• Added sprint plan under docs/sprints/sampling-swappable-store/

Testing and Quality

• Added unit coverage for injected-store construction, Protocol conformance, and sampling via injected stores
• Added ScoredPairProvider to the vulture whitelist
• make check passes fully

v0.4.0

What's New in 0.4.0

New Module: database_cache

• Local-first SQLite cache for the vocabulary practice loop — serves reads and accepts writes locally, syncs with remote Neon in the background
• Full public API: DatabaseCacheService with init(), get_words(), get_word_pairs_with_scores(), record_exercise_result(), refresh(), flush(), get_status()
• Stale-while-revalidate refresh policy — fresh data served immediately, background refresh when stale
• Transactional outbox pattern — local score writes + event recording in a single SQLite transaction
• Idempotent remote replay — event IDs prevent double-application of score deltas on flush
• Atomic snapshot rebuild — refresh replaces the full snapshot and reapplies pending events in one transaction
• 4 SQLite tables: cached_word_pairs, cached_scores, pending_score_events, cache_metadata
• Comprehensive test suite: 44 tests (28 unit, 13 integration, 3 E2E against real Neon)

Auto-Flush

• record_exercise_result() now automatically triggers a fire-and-forget background flush via asyncio.create_task
• flush() remains available as a public method for explicit use
• Background auto-flush uses skip-if-running semantics; explicit flush() waits for completion

Breaking Changes

• Removed CachedDatabaseService from database module (legacy, superseded by database_cache)

Bug Fixes

• Fixed flush race condition: explicit flush() calls now wait for any running background flush to complete instead of silently skipping. Previously, calling flush() right after record_exercise_result() could return as a no-op if the auto-flush held the lock.
• Fixed apply_score_delta to validate delta values (+1/-1) and exercise types (FR32)
• Made apply_score_delta atomic (single-transaction check+increment+mark)
• Added missing FR21 implementation: get_words() logs warning for untranslated words

Documentation

• All module docs synced with implementation — removed phantom Protocol references, fixed architecture diagrams
• Shared docs/architecture.md and docs/prd.md updated to reference database_cache and remove CachedDatabaseService
• All doc files trimmed to ≤200 lines (repo convention)
• PRD FR38-41 updated to document atomic score replay and delta validation
• Architecture doc corrected: sampling reads directly from database, not through database_cache
• Removed stale database_cached/ directories from prior module rename

Other Changes

• Added aiosqlite dependency
• Updated vulture_whitelist.py for new public API symbols
• Removed obsolete sampling integration tests and latency benchmark test

v0.3.0

What's New in 0.3.0

Breaking Changes

• ExerciseProgressStore constructor now requires exercise_types: list[str] parameter
• ExerciseProgressStore.increment() takes source_word_id: int instead of source_word: Word
• ExerciseProgressStore.get_word_pairs_with_scores() no longer accepts exercise_types parameter (uses configured set)
• ScoredWordPair model now includes source_word_id: int field

New Features

• Per-exercise-type score tables: Exercise statistics are now stored in separate database tables per exercise type (e.g., user_word_exercise_scores_nl_ru_flashcard), improving isolation and making table ownership explicit
• Cache-support APIs: New export_remote_snapshot() and apply_score_delta(event_id, ...) methods on ExerciseProgressStore for the planned database_cache module
• Idempotent score replay: apply_score_delta uses event IDs to prevent double-application of remote score deltas
• database_cache module documentation: Full PRD, architecture, and product brief for the upcoming local-first cache module

Bug Fixes

• Fixed DatabaseService.add_words() to use the word's actual language attribute for table routing instead of always defaulting to the source language

Other Changes

• CachedDatabaseService marked as legacy (superseded by planned database_cache)
• NeonBackend exercise methods extracted to _neon_exercise.py to stay under 200-line limit
• Updated README with database_cache recommendation and capabilities overview
• sampling module updated to pass exercise_types to ExerciseProgressStore

v0.2.2

• add interjection to the shared PartOfSpeech taxonomy and align translate_word / extract_words prompts
• update tests for interjection handling and rotated vocabulary expectations
• stop tracking Python cache files and ignore generated build artifacts in release checks

v0.2.1

Fix packaging of runtime prompt JSON artifacts.

• include prompt JSON files in built wheels/sdists for all prompt-backed modules
• add regression tests that verify wheel contents for extract_text_from_image, extract_words_from_text, translate_text, and translate_word
• fix WordExtractor(Language.NL) crash from missing packaged prompt file

v0.2.0

What's New

Database module (database)

• DatabaseService — persist words and translations to Neon PostgreSQL
• CachedDatabaseService — LRU-cached variant for read-heavy workloads
• ExerciseProgressStore — per-user per-exercise score tracking
• Neon backend with asyncpg, abstract backend ABC for extensibility
• Testing utilities (drop_all_tables, reset_database, helpers)

Sampling module (sampling)

• WordSampler — weighted word sampling with adversarial distractor selection

Docs

• Added DATABASE_URL to ENV_VARS.md
• Added sampling module to README
• Updated database module description in README
• NEON.md with full Neon PostgreSQL setup instructions

Tests

• 133 unit tests passing
• Integration and e2e tests for database and sampling modules