karaoke-gen 0.75.54__py3-none-any.whl

lyrics_transcriber/correction/agentic/adapter.py

@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+from typing import Dict, Any, List
+
+from .models.schemas import CorrectionProposal
+from lyrics_transcriber.types import WordCorrection, Word
+from lyrics_transcriber.utils.word_utils import WordUtils
+
+
+def adapt_proposals_to_word_corrections(
+    proposals: List[CorrectionProposal],
+    word_map: Dict[str, Word],
+    linear_position_map: Dict[str, int],
+) -> List[WordCorrection]:
+    """Convert CorrectionProposal items into WordCorrection objects.
+
+    Minimal mapping: supports ReplaceWord and DeleteWord actions with single word_id.
+    Unknown or unsupported actions are ignored.
+    
+    The reason field includes gap category and confidence for better UI feedback.
+    """
+    results: List[WordCorrection] = []
+    for p in proposals:
+        action = (p.action or "").lower()
+        target_id = p.word_id or (p.word_ids[0] if p.word_ids else None)
+        if not target_id or target_id not in word_map:
+            continue
+        original = word_map[target_id]
+        original_position = linear_position_map.get(target_id, 0)
+        
+        # Build a detailed reason including gap category
+        category_str = f" [{p.gap_category.value}]" if p.gap_category else ""
+        confidence_str = f" (confidence: {p.confidence:.0%})" if p.confidence else ""
+        detailed_reason = f"{p.reason or 'AI correction'}{category_str}{confidence_str}"
+
+        if action == "replaceword" and p.replacement_text:
+            results.append(
+                WordCorrection(
+                    original_word=original.text,
+                    corrected_word=p.replacement_text,
+                    original_position=original_position,
+                    source="agentic",
+                    reason=detailed_reason,
+                    confidence=float(p.confidence or 0.0),
+                    is_deletion=False,
+                    word_id=target_id,
+                    corrected_word_id=WordUtils.generate_id(),  # Generate unique ID for corrected word
+                    handler="AgenticCorrector",  # Required by frontend
+                    reference_positions={},  # Required by frontend
+                )
+            )
+        elif action == "deleteword":
+            results.append(
+                WordCorrection(
+                    original_word=original.text,
+                    corrected_word="",
+                    original_position=original_position,
+                    source="agentic",
+                    reason=detailed_reason,
+                    confidence=float(p.confidence or 0.0),
+                    is_deletion=True,
+                    word_id=target_id,
+                    corrected_word_id=None,  # Deleted words don't need a corrected ID
+                    handler="AgenticCorrector",  # Required by frontend
+                    reference_positions={},  # Required by frontend
+                )
+            )
+
+    return results
+
+

lyrics_transcriber/correction/agentic/agent.py

@@ -0,0 +1,313 @@
+from __future__ import annotations
+
+import logging
+import os
+import json
+from typing import Dict, Any, List, Optional
+
+from .providers.base import BaseAIProvider
+from .providers.langchain_bridge import LangChainBridge
+from .providers.config import ProviderConfig
+from .models.schemas import CorrectionProposal, GapClassification, GapCategory
+from .workflows.correction_graph import build_correction_graph
+from .prompts.classifier import build_classification_prompt
+from .handlers.registry import HandlerRegistry
+
+logger = logging.getLogger(__name__)
+
+
+class AgenticCorrector:
+    """Main entry for agentic AI correction using LangChain + LangGraph.
+
+    This orchestrates correction workflows using LangGraph for state management
+    and LangChain ChatModels for provider integration. Langfuse tracing is
+    automatic via LangChain callbacks.
+    
+    Uses dependency injection for better testability - you can inject a
+    mock provider for testing.
+    """
+
+    def __init__(
+        self, 
+        provider: BaseAIProvider,
+        graph: Optional[Any] = None,
+        langfuse_handler: Optional[Any] = None,
+        session_id: Optional[str] = None
+    ):
+        """Initialize with injected dependencies.
+        
+        Args:
+            provider: AI provider implementation (e.g., LangChainBridge)
+            graph: Optional LangGraph workflow (builds default if None)
+            langfuse_handler: Optional Langfuse callback handler (if None, will try to get from provider)
+            session_id: Optional Langfuse session ID to group related traces
+        """
+        self._provider = provider
+        self._session_id = session_id
+        
+        # Get Langfuse handler from provider if available (avoids duplication)
+        self._langfuse_handler = langfuse_handler or self._get_provider_handler()
+        
+        # Build graph with Langfuse callback if available
+        self._graph = graph if graph is not None else build_correction_graph(
+            callbacks=[self._langfuse_handler] if self._langfuse_handler else None
+        )
+    
+    def _get_provider_handler(self) -> Optional[Any]:
+        """Get Langfuse handler from provider if it has one.
+        
+        This avoids duplicating Langfuse initialization - if the provider
+        (e.g., LangChainBridge) already has a handler, we reuse it.
+        
+        Returns:
+            CallbackHandler instance from provider, or None
+        """
+        # Check if provider is LangChainBridge and has a factory
+        if hasattr(self._provider, '_factory'):
+            factory = self._provider._factory
+            
+            # Force initialization of Langfuse if keys are present
+            # This ensures the handler is available when we need it
+            if hasattr(factory, '_langfuse_initialized'):
+                if not factory._langfuse_initialized:
+                    # Initialize by calling _create_callbacks (which triggers _initialize_langfuse)
+                    factory._create_callbacks(self._provider._model)
+            
+            # Now check if handler is available
+            if hasattr(factory, '_langfuse_handler'):
+                handler = factory._langfuse_handler
+                if handler:
+                    logger.debug("🤖 Reusing Langfuse handler from ModelFactory")
+                    return handler
+        
+        logger.debug("🤖 No Langfuse handler from provider")
+        return None
+    
+    @classmethod
+    def from_model(
+        cls, 
+        model: str, 
+        config: ProviderConfig | None = None,
+        session_id: Optional[str] = None,
+        cache_dir: Optional[str] = None
+    ) -> "AgenticCorrector":
+        """Factory method to create corrector from model specification.
+        
+        This is a convenience method for the common case where you want
+        to use LangChainBridge with a model spec string.
+        
+        Args:
+            model: Model identifier in format "provider/model"
+            config: Optional provider configuration
+            session_id: Optional Langfuse session ID to group related traces
+            cache_dir: Optional cache directory (uses default if not provided)
+            
+        Returns:
+            AgenticCorrector instance with LangChainBridge provider
+        """
+        config = config or ProviderConfig.from_env(cache_dir=cache_dir)
+        provider = LangChainBridge(model=model, config=config)
+        return cls(provider=provider, session_id=session_id)
+
+    def classify_gap(
+        self,
+        gap_id: str,
+        gap_text: str,
+        preceding_words: str,
+        following_words: str,
+        reference_contexts: Dict[str, str],
+        artist: Optional[str] = None,
+        title: Optional[str] = None
+    ) -> Optional[GapClassification]:
+        """Classify a gap using the AI provider.
+        
+        Args:
+            gap_id: Unique identifier for the gap
+            gap_text: The text of the gap
+            preceding_words: Text immediately before the gap
+            following_words: Text immediately after the gap
+            reference_contexts: Dictionary of reference lyrics from each source
+            artist: Song artist name
+            title: Song title
+        
+        Returns:
+            GapClassification object or None if classification fails
+        """
+        # Build classification prompt
+        prompt = build_classification_prompt(
+            gap_text=gap_text,
+            preceding_words=preceding_words,
+            following_words=following_words,
+            reference_contexts=reference_contexts,
+            artist=artist,
+            title=title,
+            gap_id=gap_id
+        )
+        
+        # Call AI provider to get classification
+        try:
+            data = self._provider.generate_correction_proposals(
+                prompt,
+                schema=GapClassification.model_json_schema(),
+                session_id=self._session_id
+            )
+            
+            # Extract first result
+            if data and len(data) > 0:
+                item = data[0]
+                if isinstance(item, dict) and "error" not in item:
+                    classification = GapClassification.model_validate(item)
+                    logger.debug(f"🤖 Classified gap {gap_id} as {classification.category} (confidence: {classification.confidence})")
+                    return classification
+        except Exception as e:
+            logger.warning(f"🤖 Failed to classify gap {gap_id}: {e}")
+        
+        return None
+    
+    def propose_for_gap(
+        self,
+        gap_id: str,
+        gap_words: List[Dict[str, Any]],
+        preceding_words: str,
+        following_words: str,
+        reference_contexts: Dict[str, str],
+        artist: Optional[str] = None,
+        title: Optional[str] = None
+    ) -> List[CorrectionProposal]:
+        """Generate correction proposals for a gap using two-step classification workflow.
+        
+        Args:
+            gap_id: Unique identifier for the gap
+            gap_words: List of word dictionaries with id, text, start_time, end_time
+            preceding_words: Text immediately before the gap
+            following_words: Text immediately after the gap
+            reference_contexts: Dictionary of reference lyrics from each source
+            artist: Song artist name
+            title: Song title
+        
+        Returns:
+            List of CorrectionProposal objects
+        """
+        # Step 1: Classify the gap
+        gap_text = ' '.join(w.get('text', '') for w in gap_words)
+        classification = self.classify_gap(
+            gap_id=gap_id,
+            gap_text=gap_text,
+            preceding_words=preceding_words,
+            following_words=following_words,
+            reference_contexts=reference_contexts,
+            artist=artist,
+            title=title
+        )
+        
+        if not classification:
+            # Classification failed, flag for human review
+            logger.warning(f"🤖 Classification failed for gap {gap_id}, flagging for review")
+            return [CorrectionProposal(
+                word_ids=[w['id'] for w in gap_words],
+                action="Flag",
+                confidence=0.0,
+                reason="Classification failed - unable to categorize gap",
+                requires_human_review=True,
+                artist=artist,
+                title=title
+            )]
+        
+        # Step 2: Route to appropriate handler based on category
+        try:
+            handler = HandlerRegistry.get_handler(
+                category=classification.category,
+                artist=artist,
+                title=title
+            )
+            
+            proposals = handler.handle(
+                gap_id=gap_id,
+                gap_words=gap_words,
+                preceding_words=preceding_words,
+                following_words=following_words,
+                reference_contexts=reference_contexts,
+                classification_reasoning=classification.reasoning
+            )
+            
+            # Add classification metadata to proposals
+            for proposal in proposals:
+                if not proposal.gap_category:
+                    proposal.gap_category = classification.category
+                if not proposal.artist:
+                    proposal.artist = artist
+                if not proposal.title:
+                    proposal.title = title
+            
+            return proposals
+            
+        except Exception as e:
+            logger.error(f"🤖 Handler failed for gap {gap_id} (category: {classification.category}): {e}")
+            # Handler failed, flag for human review
+            return [CorrectionProposal(
+                word_ids=[w['id'] for w in gap_words],
+                action="Flag",
+                confidence=0.0,
+                reason=f"Handler error for category {classification.category}: {str(e)}",
+                gap_category=classification.category,
+                requires_human_review=True,
+                artist=artist,
+                title=title
+            )]
+
+    def propose(self, prompt: str) -> List[CorrectionProposal]:
+        """Generate correction proposals using LangGraph + LangChain.
+        
+        DEPRECATED: This method uses the old single-step approach.
+        Use propose_for_gap() for the new two-step classification workflow.
+        
+        Args:
+            prompt: The correction prompt with gap text and reference context
+            
+        Returns:
+            List of validated CorrectionProposal objects
+        """
+        # Prepare config with session_id in metadata (Langfuse format)
+        config = {}
+        if self._langfuse_handler:
+            config["callbacks"] = [self._langfuse_handler]
+            if self._session_id:
+                config["metadata"] = {"langfuse_session_id": self._session_id}
+                logger.debug(f"🤖 Set Langfuse session_id in metadata: {self._session_id}")
+        
+        # Run LangGraph workflow (with Langfuse tracing if configured)
+        if self._graph:
+            try:
+                self._graph.invoke(
+                    {"prompt": prompt, "proposals": []},
+                    config=config
+                )
+            except Exception as e:
+                logger.debug(f"🤖 LangGraph workflow invocation failed: {e}")
+
+        # Get proposals from LangChain ChatModel
+        # Pass the session_id via metadata to the provider
+        data = self._provider.generate_correction_proposals(
+            prompt, 
+            schema=CorrectionProposal.model_json_schema(),
+            session_id=self._session_id
+        )
+        
+        # Validate via Pydantic; invalid entries are dropped
+        proposals: List[CorrectionProposal] = []
+        for item in data:
+            # Check if this is an error response from the provider
+            if isinstance(item, dict) and "error" in item:
+                logger.warning(f"🤖 Provider returned error: {item}")
+                continue
+            
+            try:
+                proposals.append(CorrectionProposal.model_validate(item))
+            except Exception as e:
+                # Log validation errors for debugging
+                logger.debug(f"🤖 Failed to validate proposal: {e}, item: {item}")
+                continue
+                
+        return proposals
+
+

lyrics_transcriber/correction/agentic/feedback/aggregator.py

@@ -0,0 +1,12 @@
+from __future__ import annotations
+
+from typing import Dict, Any
+
+
+class FeedbackAggregator:
+    """Placeholder for learning data aggregation logic."""
+
+    def aggregate(self, session_id: str) -> Dict[str, Any]:
+        return {"session_id": session_id, "status": "ok"}
+
+

lyrics_transcriber/correction/agentic/feedback/collector.py

@@ -0,0 +1,17 @@
+from __future__ import annotations
+
+from typing import Dict, Any
+
+from .store import FeedbackStore
+
+
+class FeedbackCollector:
+    def __init__(self, store: FeedbackStore | None):
+        self._store = store
+
+    def collect(self, feedback_id: str, session_id: str | None, data_json: str) -> None:
+        if not self._store:
+            return
+        self._store.put_feedback(feedback_id, session_id, data_json)
+
+

lyrics_transcriber/correction/agentic/feedback/retention.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+import sqlite3
+from datetime import datetime, timedelta
+from typing import Optional
+
+
+def cleanup_expired(db_path: str, older_than_days: int = 365 * 3) -> int:
+    """Cleanup routine placeholder; returns number of deleted rows.
+
+    Note: This placeholder assumes `data` JSON contains an ISO timestamp under
+    key `createdAt`. For production, store timestamps as columns.
+    """
+    threshold = (datetime.utcnow() - timedelta(days=older_than_days)).isoformat()
+    with sqlite3.connect(db_path) as conn:
+        cur = conn.cursor()
+        # Delete sessions and feedback older than threshold by created_at
+        cur.execute("DELETE FROM sessions WHERE created_at < ?", (threshold,))
+        cur.execute("DELETE FROM feedback WHERE created_at < ?", (threshold,))
+        deleted = cur.rowcount
+        conn.commit()
+        return deleted
+
+

lyrics_transcriber/correction/agentic/feedback/store.py

@@ -0,0 +1,76 @@
+from __future__ import annotations
+
+import sqlite3
+from dataclasses import asdict
+from pathlib import Path
+from typing import Dict, Any, Iterable, Optional
+from datetime import datetime
+
+
+class FeedbackStore:
+    """SQLite-backed store for sessions, corrections, and feedback.
+
+    This is a minimal implementation to satisfy contract needs; schema may
+    evolve. All operations are simple and synchronous for local usage.
+    """
+
+    def __init__(self, db_path: str | Path):
+        self._db_path = str(db_path)
+        self._init()
+
+    def _init(self) -> None:
+        with sqlite3.connect(self._db_path) as conn:
+            cur = conn.cursor()
+            cur.execute(
+                """
+                CREATE TABLE IF NOT EXISTS sessions (
+                    id TEXT PRIMARY KEY,
+                    data TEXT NOT NULL,
+                    created_at TEXT NOT NULL
+                )
+                """
+            )
+            cur.execute(
+                """
+                CREATE TABLE IF NOT EXISTS feedback (
+                    id TEXT PRIMARY KEY,
+                    session_id TEXT,
+                    data TEXT NOT NULL,
+                    created_at TEXT NOT NULL
+                )
+                """
+            )
+            # Attempt to add created_at if upgrading from older schema
+            try:
+                cur.execute("ALTER TABLE sessions ADD COLUMN created_at TEXT")
+            except Exception:
+                pass
+            try:
+                cur.execute("ALTER TABLE feedback ADD COLUMN created_at TEXT")
+            except Exception:
+                pass
+            conn.commit()
+
+    def put_session(self, session_id: str, data_json: str) -> None:
+        with sqlite3.connect(self._db_path) as conn:
+            conn.execute(
+                "REPLACE INTO sessions (id, data, created_at) VALUES (?, ?, ?)",
+                (session_id, data_json, datetime.utcnow().isoformat()),
+            )
+            conn.commit()
+
+    def get_session(self, session_id: str) -> Optional[str]:
+        with sqlite3.connect(self._db_path) as conn:
+            cur = conn.execute("SELECT data FROM sessions WHERE id = ?", (session_id,))
+            row = cur.fetchone()
+            return row[0] if row else None
+
+    def put_feedback(self, feedback_id: str, session_id: Optional[str], data_json: str) -> None:
+        with sqlite3.connect(self._db_path) as conn:
+            conn.execute(
+                "REPLACE INTO feedback (id, session_id, data, created_at) VALUES (?, ?, ?, ?)",
+                (feedback_id, session_id, data_json, datetime.utcnow().isoformat()),
+            )
+            conn.commit()
+
+

lyrics_transcriber/correction/agentic/handlers/__init__.py

@@ -0,0 +1,24 @@
+"""Category-specific handlers for gap correction."""
+
+from .base import BaseHandler
+from .punctuation import PunctuationHandler
+from .sound_alike import SoundAlikeHandler
+from .background_vocals import BackgroundVocalsHandler
+from .extra_words import ExtraWordsHandler
+from .repeated_section import RepeatedSectionHandler
+from .complex_multi_error import ComplexMultiErrorHandler
+from .ambiguous import AmbiguousHandler
+from .no_error import NoErrorHandler
+
+__all__ = [
+    'BaseHandler',
+    'PunctuationHandler',
+    'SoundAlikeHandler',
+    'BackgroundVocalsHandler',
+    'ExtraWordsHandler',
+    'RepeatedSectionHandler',
+    'ComplexMultiErrorHandler',
+    'AmbiguousHandler',
+    'NoErrorHandler',
+]
+

lyrics_transcriber/correction/agentic/handlers/ambiguous.py

@@ -0,0 +1,44 @@
+"""Handler for ambiguous gaps that need human review."""
+
+from typing import List, Dict, Any
+from .base import BaseHandler
+from ..models.schemas import CorrectionProposal, GapCategory
+
+
+class AmbiguousHandler(BaseHandler):
+    """Handles ambiguous gaps where correct action is unclear without audio."""
+    
+    @property
+    def category(self) -> GapCategory:
+        return GapCategory.AMBIGUOUS
+    
+    def handle(
+        self,
+        gap_id: str,
+        gap_words: List[Dict[str, Any]],
+        preceding_words: str,
+        following_words: str,
+        reference_contexts: Dict[str, str],
+        classification_reasoning: str = ""
+    ) -> List[CorrectionProposal]:
+        """Flag ambiguous gaps for human review."""
+        
+        if not gap_words:
+            return []
+        
+        # Ambiguous cases always require human review with audio
+        gap_text = ' '.join(w.get('text', '') for w in gap_words)
+        
+        proposal = CorrectionProposal(
+            word_ids=[w['id'] for w in gap_words],
+            action="Flag",
+            confidence=0.4,
+            reason=f"Ambiguous gap: '{gap_text[:100]}...'. Cannot determine correct action without listening to audio. {classification_reasoning}",
+            gap_category=self.category,
+            requires_human_review=True,
+            artist=self.artist,
+            title=self.title
+        )
+        
+        return [proposal]
+

lyrics_transcriber/correction/agentic/handlers/background_vocals.py

@@ -0,0 +1,68 @@
+"""Handler for background vocals that should be removed."""
+
+from typing import List, Dict, Any
+from .base import BaseHandler
+from ..models.schemas import CorrectionProposal, GapCategory
+
+
+class BackgroundVocalsHandler(BaseHandler):
+    """Handles gaps containing background vocals (usually in parentheses)."""
+    
+    @property
+    def category(self) -> GapCategory:
+        return GapCategory.BACKGROUND_VOCALS
+    
+    def handle(
+        self,
+        gap_id: str,
+        gap_words: List[Dict[str, Any]],
+        preceding_words: str,
+        following_words: str,
+        reference_contexts: Dict[str, str],
+        classification_reasoning: str = ""
+    ) -> List[CorrectionProposal]:
+        """Propose deletion of words in parentheses."""
+        
+        if not gap_words:
+            return []
+        
+        proposals = []
+        
+        # Find words that are in parentheses or are parentheses themselves
+        words_to_delete = []
+        for word in gap_words:
+            text = word.get('text', '')
+            # Check if word has parentheses or is just parentheses
+            if '(' in text or ')' in text:
+                words_to_delete.append(word)
+        
+        if words_to_delete:
+            # Create delete proposals for parenthesized content
+            proposal = CorrectionProposal(
+                word_ids=[w['id'] for w in words_to_delete],
+                action="DeleteWord",
+                confidence=0.85,
+                reason=f"Background vocals in parentheses, not in reference lyrics. {classification_reasoning}",
+                gap_category=self.category,
+                requires_human_review=False,
+                artist=self.artist,
+                title=self.title
+            )
+            proposals.append(proposal)
+        else:
+            # If no parentheses found but classified as background vocals,
+            # flag for review as classifier may have other reasoning
+            proposal = CorrectionProposal(
+                word_ids=[w['id'] for w in gap_words],
+                action="Flag",
+                confidence=0.6,
+                reason=f"Classified as background vocals but no parentheses found. {classification_reasoning}",
+                gap_category=self.category,
+                requires_human_review=True,
+                artist=self.artist,
+                title=self.title
+            )
+            proposals.append(proposal)
+        
+        return proposals
+