dao-ai 0.1.2__py3-none-any.whl → 0.1.20__py3-none-any.whl

dao_ai/genie/cache/context_aware/base.py

@@ -0,0 +1,1151 @@
+"""
+Abstract base class for context-aware Genie cache implementations.
+
+This module provides the foundational abstract base class for all context-aware
+cache implementations. It extracts common code for:
+- Dual embedding generation (question + conversation context)
+- Ask question flow with error handling and graceful fallback
+- SQL execution with retry logic
+- Common properties and initialization patterns
+
+Subclasses must implement storage-specific methods:
+- _find_similar(): Find semantically similar cached entry
+- _store_entry(): Store new cache entry
+- _setup(): Initialize resources (embeddings, storage)
+- invalidate_expired(): Remove expired entries
+- clear(): Clear all entries for space
+- stats(): Return cache statistics
+"""
+
+from __future__ import annotations
+
+from abc import abstractmethod
+from datetime import timedelta
+from typing import Any, Self, TypeVar
+
+import mlflow
+import pandas as pd
+from databricks.sdk import WorkspaceClient
+from databricks.sdk.service.dashboards import (
+    GenieFeedbackRating,
+    GenieListConversationMessagesResponse,
+    GenieMessage,
+)
+from databricks_ai_bridge.genie import GenieResponse
+from loguru import logger
+
+from dao_ai.config import LLMModel, WarehouseModel
+from dao_ai.genie.cache.base import (
+    CacheResult,
+    GenieServiceBase,
+    SQLCacheEntry,
+)
+from dao_ai.genie.cache.core import execute_sql_via_warehouse
+
+# Type variable for subclass return types
+T = TypeVar("T", bound="ContextAwareGenieService")
+
+
+def get_conversation_history(
+    workspace_client: WorkspaceClient,
+    space_id: str,
+    conversation_id: str,
+    max_messages: int = 10,
+) -> list[GenieMessage]:
+    """
+    Retrieve conversation history from Genie.
+
+    Args:
+        workspace_client: The Databricks workspace client
+        space_id: The Genie space ID
+        conversation_id: The conversation ID to retrieve
+        max_messages: Maximum number of messages to retrieve
+
+    Returns:
+        List of GenieMessage objects representing the conversation history
+    """
+    try:
+        # Use the Genie API to retrieve conversation messages
+        response: GenieListConversationMessagesResponse = (
+            workspace_client.genie.list_conversation_messages(
+                space_id=space_id,
+                conversation_id=conversation_id,
+            )
+        )
+
+        # Return the most recent messages up to max_messages
+        if response.messages is not None:
+            all_messages: list[GenieMessage] = list(response.messages)
+            return (
+                all_messages[-max_messages:]
+                if len(all_messages) > max_messages
+                else all_messages
+            )
+        return []
+    except Exception as e:
+        logger.warning(
+            "Failed to retrieve conversation history",
+            conversation_id=conversation_id,
+            error=str(e),
+        )
+        return []
+
+
+def build_context_string(
+    question: str,
+    conversation_messages: list[GenieMessage],
+    window_size: int,
+    max_tokens: int = 2000,
+) -> str:
+    """
+    Build a context-aware question string using rolling window.
+
+    This function creates a concatenated string that includes recent conversation
+    turns to provide context for semantic similarity matching.
+
+    Args:
+        question: The current question
+        conversation_messages: List of previous conversation messages
+        window_size: Number of previous turns to include
+        max_tokens: Maximum estimated tokens (rough approximation: 4 chars = 1 token)
+
+    Returns:
+        Context-aware question string formatted for embedding
+    """
+    if window_size <= 0 or not conversation_messages:
+        return question
+
+    # Take the last window_size messages (most recent)
+    recent_messages = (
+        conversation_messages[-window_size:]
+        if len(conversation_messages) > window_size
+        else conversation_messages
+    )
+
+    # Build context parts
+    context_parts: list[str] = []
+
+    for msg in recent_messages:
+        # Only include messages with content from the history
+        if msg.content:
+            # Limit message length to prevent token overflow
+            content: str = msg.content
+            if len(content) > 500:  # Truncate very long messages
+                content = content[:500] + "..."
+            context_parts.append(f"Previous: {content}")
+
+    # Add current question
+    context_parts.append(f"Current: {question}")
+
+    # Join with newlines
+    context_string = "\n".join(context_parts)
+
+    # Rough token limit check (4 chars ≈ 1 token)
+    estimated_tokens = len(context_string) / 4
+    if estimated_tokens > max_tokens:
+        # Truncate to fit max_tokens
+        target_chars = max_tokens * 4
+        original_length = len(context_string)
+        context_string = context_string[:target_chars] + "..."
+        logger.trace(
+            "Truncated context string",
+            original_chars=original_length,
+            target_chars=target_chars,
+            max_tokens=max_tokens,
+        )
+
+    return context_string
+
+
+class ContextAwareGenieService(GenieServiceBase):
+    """
+    Abstract base class for context-aware Genie cache implementations.
+
+    This class provides shared implementation for:
+    - Dual embedding generation (question + conversation context)
+    - Main ask_question flow with error handling
+    - SQL execution with warehouse
+    - Common properties (time_to_live, similarity_threshold, etc.)
+
+    Subclasses must implement storage-specific methods for finding similar
+    entries, storing new entries, and managing cache lifecycle.
+
+    Error Handling:
+        All cache operations are wrapped in try/except to ensure graceful
+        degradation. If any cache operation fails, the request is delegated
+        to the underlying service without caching.
+
+    Thread Safety:
+        Subclasses are responsible for thread safety of storage operations.
+        This base class does not provide synchronization primitives.
+    """
+
+    # Common attributes - subclasses should define these
+    impl: GenieServiceBase
+    _workspace_client: WorkspaceClient | None
+    name: str
+    _embeddings: Any  # DatabricksEmbeddings
+    _embedding_dims: int | None
+    _setup_complete: bool
+
+    # Abstract methods that subclasses must implement
+    @abstractmethod
+    def _setup(self) -> None:
+        """
+        Initialize resources required by the cache implementation.
+
+        This method is called lazily before first use. Implementations should:
+        - Initialize embedding model
+        - Set up storage (database connection, in-memory structures, etc.)
+        - Create necessary tables/indexes if applicable
+
+        This method should be idempotent (safe to call multiple times).
+        """
+        pass
+
+    @abstractmethod
+    def _find_similar(
+        self,
+        question: str,
+        conversation_context: str,
+        question_embedding: list[float],
+        context_embedding: list[float],
+        conversation_id: str | None = None,
+    ) -> tuple[SQLCacheEntry, float] | None:
+        """
+        Find a semantically similar cached entry using dual embedding matching.
+
+        Args:
+            question: The original question (for logging)
+            conversation_context: The conversation context string
+            question_embedding: The embedding vector of just the question
+            context_embedding: The embedding vector of the conversation context
+            conversation_id: Optional conversation ID (for logging)
+
+        Returns:
+            Tuple of (SQLCacheEntry, combined_similarity_score) if found, None otherwise
+        """
+        pass
+
+    @abstractmethod
+    def _store_entry(
+        self,
+        question: str,
+        conversation_context: str,
+        question_embedding: list[float],
+        context_embedding: list[float],
+        response: GenieResponse,
+        message_id: str | None = None,
+    ) -> None:
+        """
+        Store a new cache entry with dual embeddings.
+
+        Args:
+            question: The user's question
+            conversation_context: Previous conversation context string
+            question_embedding: Embedding of the question
+            context_embedding: Embedding of the conversation context
+            response: The GenieResponse containing query, description, etc.
+            message_id: The Genie message ID from the original API response.
+                Stored with the cache entry to enable feedback on cache hits.
+        """
+        pass
+
+    def invalidate_expired(self) -> int | dict[str, int]:
+        """
+        Template method for removing expired entries from the cache.
+
+        This method implements the TTL check and delegates to
+        _delete_expired_entries() for the actual deletion.
+
+        Returns:
+            Number of entries deleted, or dict with counts by category
+        """
+        self._setup()
+        ttl_seconds = self.time_to_live_seconds
+
+        if ttl_seconds is None or ttl_seconds < 0:
+            return self._get_empty_expiration_result()
+
+        return self._delete_expired_entries(ttl_seconds)
+
+    @abstractmethod
+    def _delete_expired_entries(self, ttl_seconds: int) -> int | dict[str, int]:
+        """
+        Delete expired entries from storage.
+
+        Args:
+            ttl_seconds: TTL in seconds for determining expiration
+
+        Returns:
+            Number of entries deleted, or dict with counts by category
+        """
+        pass
+
+    def _get_empty_expiration_result(self) -> int | dict[str, int]:
+        """
+        Return the empty result for invalidate_expired when TTL is disabled.
+
+        Override this in subclasses that return dict to return appropriate empty dict.
+
+        Returns:
+            0 by default, or empty dict for subclasses that return dict
+        """
+        return 0
+
+    def clear(self) -> int:
+        """
+        Template method for clearing all entries from the cache.
+
+        This method calls _setup() and delegates to _delete_all_entries().
+
+        Returns:
+            Number of entries deleted
+        """
+        self._setup()
+        return self._delete_all_entries()
+
+    @abstractmethod
+    def _delete_all_entries(self) -> int:
+        """
+        Delete all entries for this Genie space from storage.
+
+        Returns:
+            Number of entries deleted
+        """
+        pass
+
+    def stats(self) -> dict[str, Any]:
+        """
+        Template method for returning cache statistics.
+
+        This method uses the Template Method pattern to consolidate the common
+        stats calculation algorithm. Subclasses provide counting implementations
+        via abstract methods and can add additional stats via hook methods.
+
+        Returns:
+            Dict with cache statistics (size, ttl, thresholds, etc.)
+        """
+        self._setup()
+        ttl_seconds = self.time_to_live_seconds
+        ttl = self.time_to_live
+
+        # Calculate base stats using abstract counting methods
+        if ttl_seconds is None or ttl_seconds < 0:
+            total = self._count_all_entries()
+            base_stats: dict[str, Any] = {
+                "size": total,
+                "ttl_seconds": None,
+                "similarity_threshold": self.similarity_threshold,
+                "expired_entries": 0,
+                "valid_entries": total,
+            }
+        else:
+            total, expired = self._count_entries_with_ttl(ttl_seconds)
+            base_stats = {
+                "size": total,
+                "ttl_seconds": ttl.total_seconds() if ttl else None,
+                "similarity_threshold": self.similarity_threshold,
+                "expired_entries": expired,
+                "valid_entries": total - expired,
+            }
+
+        # Add any additional stats from subclasses
+        additional_stats = self._get_additional_stats()
+        base_stats.update(additional_stats)
+
+        return base_stats
+
+    @abstractmethod
+    def _count_all_entries(self) -> int:
+        """
+        Count all cache entries for this Genie space.
+
+        Returns:
+            Total number of cache entries
+        """
+        pass
+
+    @abstractmethod
+    def _count_entries_with_ttl(self, ttl_seconds: int) -> tuple[int, int]:
+        """
+        Count total and expired entries for this Genie space.
+
+        Args:
+            ttl_seconds: TTL in seconds for determining expiration
+
+        Returns:
+            Tuple of (total_entries, expired_entries)
+        """
+        pass
+
+    def _get_additional_stats(self) -> dict[str, Any]:
+        """
+        Hook method to add additional stats from subclasses.
+
+        Override this method to add subclass-specific statistics like
+        capacity (in-memory) or prompt history stats (postgres).
+
+        Returns:
+            Dict with additional stats to merge into base stats
+        """
+        return {}
+
+    # Properties that subclasses should implement or inherit
+    @property
+    @abstractmethod
+    def warehouse(self) -> WarehouseModel:
+        """The warehouse used for executing cached SQL queries."""
+        pass
+
+    @property
+    @abstractmethod
+    def time_to_live(self) -> timedelta | None:
+        """Time-to-live for cache entries. None means never expires."""
+        pass
+
+    @property
+    @abstractmethod
+    def similarity_threshold(self) -> float:
+        """Minimum similarity for cache hit (using L2 distance converted to similarity)."""
+        pass
+
+    @property
+    def embedding_dims(self) -> int:
+        """Dimension size for embeddings (auto-detected if not configured)."""
+        if self._embedding_dims is None:
+            raise RuntimeError(
+                "Embedding dimensions not yet initialized. Call _setup() first."
+            )
+        return self._embedding_dims
+
+    @property
+    def space_id(self) -> str:
+        """The Genie space ID from the underlying service."""
+        return self.impl.space_id
+
+    @property
+    def workspace_client(self) -> WorkspaceClient | None:
+        """Get workspace client, delegating to impl if not set."""
+        if self._workspace_client is not None:
+            return self._workspace_client
+        return self.impl.workspace_client
+
+    @property
+    def time_to_live_seconds(self) -> int | None:
+        """TTL in seconds (None or negative = never expires)."""
+        ttl = self.time_to_live
+        if ttl is None:
+            return None
+        return int(ttl.total_seconds())
+
+    # Abstract method for embedding - subclasses must implement
+    @abstractmethod
+    def _embed_question(
+        self, question: str, conversation_id: str | None = None
+    ) -> tuple[list[float], list[float], str]:
+        """
+        Generate dual embeddings for a question with conversation context.
+
+        Args:
+            question: The question to embed
+            conversation_id: Optional conversation ID for retrieving context
+
+        Returns:
+            Tuple of (question_embedding, context_embedding, conversation_context_string)
+        """
+        pass
+
+    # Shared implementation methods
+    def initialize(self) -> Self:
+        """
+        Eagerly initialize the cache service.
+
+        Call this during tool creation to:
+        - Validate configuration early (fail fast)
+        - Initialize resources before any requests
+        - Avoid first-request latency from lazy initialization
+
+        Returns:
+            self for method chaining
+        """
+        self._setup()
+        return self
+
+    def _initialize_embeddings(
+        self,
+        embedding_model: str | LLMModel,
+        embedding_dims: int | None = None,
+    ) -> None:
+        """
+        Initialize the embeddings model and detect dimensions.
+
+        This helper method handles embedding model initialization for subclasses.
+
+        Args:
+            embedding_model: The embedding model name or LLMModel instance
+            embedding_dims: Optional pre-configured embedding dimensions
+        """
+        # Convert embedding_model to LLMModel if it's a string
+        model: LLMModel = (
+            LLMModel(name=embedding_model)
+            if isinstance(embedding_model, str)
+            else embedding_model
+        )
+        self._embeddings = model.as_embeddings_model()
+
+        # Auto-detect embedding dimensions if not provided
+        if embedding_dims is None:
+            sample_embedding: list[float] = self._embeddings.embed_query("test")
+            self._embedding_dims = len(sample_embedding)
+            logger.debug(
+                "Auto-detected embedding dimensions",
+                layer=self.name,
+                dims=self._embedding_dims,
+            )
+        else:
+            self._embedding_dims = embedding_dims
+
+    def _embed_question_with_genie_history(
+        self,
+        question: str,
+        conversation_id: str | None,
+        context_window_size: int,
+        max_context_tokens: int,
+    ) -> tuple[list[float], list[float], str]:
+        """
+        Generate dual embeddings using Genie API for conversation history.
+
+        This method retrieves conversation history from the Genie API and
+        generates dual embeddings for semantic matching.
+
+        Args:
+            question: The question to embed
+            conversation_id: Optional conversation ID for retrieving context
+            context_window_size: Number of previous messages to include
+            max_context_tokens: Maximum tokens for context string
+
+        Returns:
+            Tuple of (question_embedding, context_embedding, conversation_context_string)
+        """
+        conversation_context = ""
+
+        # If conversation context is enabled and available
+        if (
+            self.workspace_client is not None
+            and conversation_id is not None
+            and context_window_size > 0
+        ):
+            try:
+                # Retrieve conversation history from Genie API
+                conversation_messages = get_conversation_history(
+                    workspace_client=self.workspace_client,
+                    space_id=self.space_id,
+                    conversation_id=conversation_id,
+                    max_messages=context_window_size * 2,  # Get extra for safety
+                )
+
+                # Build context string
+                if conversation_messages:
+                    recent_messages = (
+                        conversation_messages[-context_window_size:]
+                        if len(conversation_messages) > context_window_size
+                        else conversation_messages
+                    )
+
+                    context_parts: list[str] = []
+                    for msg in recent_messages:
+                        if msg.content:
+                            content: str = msg.content
+                            if len(content) > 500:
+                                content = content[:500] + "..."
+                            context_parts.append(f"Previous: {content}")
+
+                    conversation_context = "\n".join(context_parts)
+
+                    # Truncate if too long
+                    estimated_tokens = len(conversation_context) / 4
+                    if estimated_tokens > max_context_tokens:
+                        target_chars = max_context_tokens * 4
+                        conversation_context = (
+                            conversation_context[:target_chars] + "..."
+                        )
+
+                logger.trace(
+                    "Using conversation context from Genie API",
+                    layer=self.name,
+                    messages_count=len(conversation_messages),
+                    window_size=context_window_size,
+                )
+            except Exception as e:
+                logger.warning(
+                    "Failed to build conversation context, using question only",
+                    layer=self.name,
+                    error=str(e),
+                )
+                conversation_context = ""
+
+        return self._generate_dual_embeddings(question, conversation_context)
+
+    def _generate_dual_embeddings(
+        self, question: str, conversation_context: str
+    ) -> tuple[list[float], list[float], str]:
+        """
+        Generate dual embeddings for question and conversation context.
+
+        Args:
+            question: The question to embed
+            conversation_context: The conversation context string
+
+        Returns:
+            Tuple of (question_embedding, context_embedding, conversation_context)
+        """
+        if conversation_context:
+            # Embed both question and context
+            embeddings: list[list[float]] = self._embeddings.embed_documents(
+                [question, conversation_context]
+            )
+            question_embedding = embeddings[0]
+            context_embedding = embeddings[1]
+        else:
+            # Only embed question, use zero vector for context
+            embeddings = self._embeddings.embed_documents([question])
+            question_embedding = embeddings[0]
+            context_embedding = [0.0] * len(question_embedding)  # Zero vector
+
+        return question_embedding, context_embedding, conversation_context
+
+    @mlflow.trace(name="execute_cached_sql")
+    def _execute_sql(self, sql: str) -> pd.DataFrame | str:
+        """
+        Execute SQL using the warehouse and return results.
+
+        Args:
+            sql: The SQL query to execute
+
+        Returns:
+            DataFrame with results, or error message string if execution failed
+        """
+        return execute_sql_via_warehouse(
+            warehouse=self.warehouse,
+            sql=sql,
+            layer_name=self.name,
+        )
+
+    def _build_cache_hit_response(
+        self,
+        cached: SQLCacheEntry,
+        result: pd.DataFrame,
+        conversation_id: str | None,
+    ) -> CacheResult:
+        """
+        Build a CacheResult for a cache hit.
+
+        Args:
+            cached: The cached SQL entry
+            result: The fresh DataFrame from SQL execution
+            conversation_id: The current conversation ID
+
+        Returns:
+            CacheResult with cache_hit=True, including message_id and cache_entry_id
+            from the original cached entry for traceability and feedback support.
+        """
+        # IMPORTANT: Use the current conversation_id (from the request), not the cached one
+        # This ensures the conversation continues properly
+        response = GenieResponse(
+            result=result,
+            query=cached.query,
+            description=cached.description,
+            conversation_id=conversation_id
+            if conversation_id
+            else cached.conversation_id,
+        )
+        # Cache hit - include message_id from original response for feedback support
+        # and cache_entry_id for traceability to genie_prompt_history
+        return CacheResult(
+            response=response,
+            cache_hit=True,
+            served_by=self.name,
+            message_id=cached.message_id,
+            cache_entry_id=cached.cache_entry_id,
+        )
+
+    def ask_question(
+        self, question: str, conversation_id: str | None = None
+    ) -> CacheResult:
+        """
+        Ask a question, using semantic cache if a similar query exists.
+
+        On cache hit, re-executes the cached SQL to get fresh data.
+        Returns CacheResult with cache metadata.
+
+        This method wraps ask_question_with_cache_info with error handling
+        to ensure graceful degradation on cache failures.
+
+        Args:
+            question: The question to ask
+            conversation_id: Optional conversation ID for context
+
+        Returns:
+            CacheResult with fresh response and cache metadata
+        """
+        try:
+            return self.ask_question_with_cache_info(question, conversation_id)
+        except Exception as e:
+            logger.warning(
+                "Context-aware cache operation failed, delegating to underlying service",
+                layer=self.name,
+                error=str(e),
+                exc_info=True,
+            )
+            # Graceful degradation: fall back to underlying service
+            return self.impl.ask_question(question, conversation_id)
+
+    def ask_question_with_cache_info(
+        self,
+        question: str,
+        conversation_id: str | None = None,
+    ) -> CacheResult:
+        """
+        Template method for asking a question with cache lookup.
+
+        This method implements the cache lookup algorithm using the Template Method
+        pattern. Subclasses can customize behavior by overriding hook methods:
+        - _before_cache_lookup(): Called before cache search (e.g., store prompt)
+        - _after_cache_hit(): Called after a cache hit (e.g., update prompt flags)
+        - _after_cache_miss(): Called after a cache miss (e.g., store prompt)
+
+        Args:
+            question: The question to ask
+            conversation_id: Optional conversation ID for context and continuation
+
+        Returns:
+            CacheResult with fresh response and cache metadata
+        """
+        self._setup()
+
+        # Step 1: Generate dual embeddings
+        question_embedding, context_embedding, conversation_context = (
+            self._embed_question(question, conversation_id)
+        )
+
+        # Step 2: Hook for pre-lookup actions (e.g., store prompt in history)
+        self._before_cache_lookup(question, conversation_id)
+
+        # Step 3: Search for similar cached entry
+        cache_result = self._find_similar(
+            question,
+            conversation_context,
+            question_embedding,
+            context_embedding,
+            conversation_id,
+        )
+
+        # Step 4: Handle cache hit or miss
+        if cache_result is not None:
+            cached, combined_similarity = cache_result
+
+            result = self._handle_cache_hit(
+                question,
+                conversation_id,
+                cached,
+                combined_similarity,
+                conversation_context,
+                question_embedding,
+                context_embedding,
+            )
+
+            # Hook for post-cache-hit actions (e.g., update prompt cache_hit flag)
+            self._after_cache_hit(question, conversation_id, result)
+
+            return result
+
+        # Handle cache miss
+        result = self._handle_cache_miss(
+            question,
+            conversation_id,
+            conversation_context,
+            question_embedding,
+            context_embedding,
+        )
+
+        # Hook for post-cache-miss actions (e.g., store prompt if not done earlier)
+        self._after_cache_miss(question, conversation_id, result)
+
+        return result
+
+    def _before_cache_lookup(self, question: str, conversation_id: str | None) -> None:
+        """
+        Hook method called before cache lookup.
+
+        Override this method to perform actions before searching the cache,
+        such as storing the prompt in history.
+
+        Args:
+            question: The question being asked
+            conversation_id: Optional conversation ID
+        """
+        pass
+
+    def _after_cache_hit(
+        self,
+        question: str,
+        conversation_id: str | None,
+        result: CacheResult,
+    ) -> None:
+        """
+        Hook method called after a cache hit.
+
+        Override this method to perform actions after a successful cache hit,
+        such as updating prompt history flags.
+
+        Args:
+            question: The question that was asked
+            conversation_id: Optional conversation ID
+            result: The cache result
+        """
+        pass
+
+    def _after_cache_miss(
+        self,
+        question: str,
+        conversation_id: str | None,
+        result: CacheResult,
+    ) -> None:
+        """
+        Hook method called after a cache miss.
+
+        Override this method to perform actions after a cache miss,
+        such as storing prompt history if not done earlier.
+
+        Args:
+            question: The question that was asked
+            conversation_id: Optional conversation ID
+            result: The cache result
+        """
+        pass
+
+    def _handle_cache_hit(
+        self,
+        question: str,
+        conversation_id: str | None,
+        cached: SQLCacheEntry,
+        combined_similarity: float,
+        conversation_context: str,
+        question_embedding: list[float],
+        context_embedding: list[float],
+    ) -> CacheResult:
+        """
+        Handle a cache hit - execute cached SQL and return response.
+
+        This method handles the common cache hit logic including SQL execution,
+        stale cache fallback, and response building.
+
+        Args:
+            question: The original question
+            conversation_id: The conversation ID
+            cached: The cached SQL entry
+            combined_similarity: The similarity score
+            conversation_context: The conversation context string
+            question_embedding: The question embedding
+            context_embedding: The context embedding
+
+        Returns:
+            CacheResult with the response
+        """
+        logger.debug(
+            "Cache hit",
+            layer=self.name,
+            combined_similarity=f"{combined_similarity:.3f}",
+            question=question[:50],
+            conversation_id=conversation_id,
+        )
+
+        # Re-execute the cached SQL to get fresh data
+        result: pd.DataFrame | str = self._execute_sql(cached.query)
+
+        # Check if SQL execution failed (returns error string instead of DataFrame)
+        if isinstance(result, str):
+            logger.warning(
+                "Cached SQL execution failed, falling back to Genie",
+                layer=self.name,
+                question=question[:80],
+                conversation_id=conversation_id,
+                cached_sql=cached.query[:80],
+                error=result[:200],
+                space_id=self.space_id,
+            )
+
+            # Subclass should handle stale entry cleanup
+            self._on_stale_cache_entry(question)
+
+            # Fall back to Genie to get fresh SQL
+            logger.info(
+                "Delegating to Genie for fresh SQL",
+                layer=self.name,
+                question=question[:80],
+                conversation_id=conversation_id,
+                space_id=self.space_id,
+                delegating_to=type(self.impl).__name__,
+            )
+            fallback_result: CacheResult = self.impl.ask_question(
+                question, conversation_id
+            )
+
+            # Store the fresh SQL in cache
+            if fallback_result.response.query:
+                self._store_entry(
+                    question,
+                    conversation_context,
+                    question_embedding,
+                    context_embedding,
+                    fallback_result.response,
+                    message_id=fallback_result.message_id,
+                )
+                logger.info(
+                    "Stored fresh SQL from fallback",
+                    layer=self.name,
+                    fresh_sql=fallback_result.response.query[:80],
+                    space_id=self.space_id,
+                    message_id=fallback_result.message_id,
+                )
+            else:
+                logger.warning(
+                    "Fallback response has no SQL query to cache",
+                    layer=self.name,
+                    question=question[:80],
+                    space_id=self.space_id,
+                )
+
+            # Return as cache miss (fallback scenario)
+            # Propagate message_id from fallback result
+            return CacheResult(
+                response=fallback_result.response,
+                cache_hit=False,
+                served_by=None,
+                message_id=fallback_result.message_id,
+            )
+
+        # Build and return cache hit response
+        return self._build_cache_hit_response(cached, result, conversation_id)
+
+    def _on_stale_cache_entry(self, question: str) -> None:
+        """
+        Called when a stale cache entry is detected (SQL execution failed).
+
+        Subclasses can override this to clean up the stale entry from storage.
+
+        Args:
+            question: The question that had a stale cache entry
+        """
+        # Default implementation does nothing - subclasses should override
+        pass
+
+    def _handle_cache_miss(
+        self,
+        question: str,
+        conversation_id: str | None,
+        conversation_context: str,
+        question_embedding: list[float],
+        context_embedding: list[float],
+    ) -> CacheResult:
+        """
+        Handle a cache miss - delegate to underlying service and store result.
+
+        Args:
+            question: The original question
+            conversation_id: The conversation ID
+            conversation_context: The conversation context string
+            question_embedding: The question embedding
+            context_embedding: The context embedding
+
+        Returns:
+            CacheResult from the underlying service
+        """
+        logger.info(
+            "Cache MISS",
+            layer=self.name,
+            question=question[:80],
+            conversation_id=conversation_id,
+            space_id=self.space_id,
+            similarity_threshold=self.similarity_threshold,
+            delegating_to=type(self.impl).__name__,
+        )
+
+        result: CacheResult = self.impl.ask_question(question, conversation_id)
+
+        # Store in cache if we got a SQL query
+        if result.response.query:
+            logger.debug(
+                "Storing new cache entry",
+                layer=self.name,
+                question=question[:50],
+                conversation_id=conversation_id,
+                space=self.space_id,
+                message_id=result.message_id,
+            )
+            self._store_entry(
+                question,
+                conversation_context,
+                question_embedding,
+                context_embedding,
+                result.response,
+                message_id=result.message_id,
+            )
+        else:
+            logger.warning(
+                "Not caching: response has no SQL query",
+                layer=self.name,
+                question=question[:50],
+            )
+
+        # Propagate message_id from underlying service result
+        return CacheResult(
+            response=result.response,
+            cache_hit=False,
+            served_by=None,
+            message_id=result.message_id,
+        )
+
+    @abstractmethod
+    def _invalidate_by_question(self, question: str) -> bool:
+        """
+        Invalidate cache entries matching a specific question.
+
+        This method is called when negative feedback is received to remove
+        the corresponding cache entry.
+
+        Args:
+            question: The question text to match and invalidate
+
+        Returns:
+            True if an entry was found and invalidated, False otherwise
+        """
+        pass
+
+    @mlflow.trace(name="genie_context_aware_send_feedback")
+    def send_feedback(
+        self,
+        conversation_id: str,
+        rating: GenieFeedbackRating,
+        message_id: str | None = None,
+        was_cache_hit: bool = False,
+    ) -> None:
+        """
+        Send feedback for a Genie message with cache invalidation.
+
+        For context-aware caches, this method:
+        1. If was_cache_hit is False: forwards feedback to the underlying service
+        2. If rating is NEGATIVE: invalidates any matching cache entries
+
+        Args:
+            conversation_id: The conversation containing the message
+            rating: The feedback rating (POSITIVE, NEGATIVE, or NONE)
+            message_id: Optional message ID. If None, looks up the most recent message.
+            was_cache_hit: Whether the response being rated was served from cache.
+
+        Note:
+            For cached responses (was_cache_hit=True), only cache invalidation is
+            performed. No feedback is sent to the Genie API because cached responses
+            don't have a corresponding Genie message.
+
+            Future Enhancement: To enable full Genie feedback for cached responses,
+            the cache would need to store the original message_id. This would require:
+            1. Adding message_id column to cache tables
+            2. Adding message_id field to SQLCacheEntry dataclass
+            3. Capturing message_id from the original Genie API response
+               (databricks_ai_bridge.genie.GenieResponse doesn't expose this)
+            4. Using WorkspaceClient directly instead of databricks_ai_bridge
+        """
+        invalidated = False
+
+        # Handle cache invalidation on negative feedback
+        if rating == GenieFeedbackRating.NEGATIVE:
+            # Need to look up the message content to find matching cache entries
+            if self.workspace_client is not None:
+                from dao_ai.genie.cache.base import (
+                    get_latest_message_id,
+                    get_message_content,
+                )
+
+                # Get message_id if not provided
+                target_message_id = message_id
+                if target_message_id is None:
+                    target_message_id = get_latest_message_id(
+                        workspace_client=self.workspace_client,
+                        space_id=self.space_id,
+                        conversation_id=conversation_id,
+                    )
+
+                # Get the message content (question) to find matching cache entries
+                if target_message_id:
+                    question = get_message_content(
+                        workspace_client=self.workspace_client,
+                        space_id=self.space_id,
+                        conversation_id=conversation_id,
+                        message_id=target_message_id,
+                    )
+
+                    if question:
+                        invalidated = self._invalidate_by_question(question)
+                        if invalidated:
+                            logger.info(
+                                "Invalidated cache entry due to negative feedback",
+                                layer=self.name,
+                                question=question[:80],
+                                conversation_id=conversation_id,
+                                message_id=target_message_id,
+                            )
+                        else:
+                            logger.debug(
+                                "No cache entry found to invalidate for negative feedback",
+                                layer=self.name,
+                                question=question[:80],
+                                conversation_id=conversation_id,
+                            )
+                    else:
+                        logger.warning(
+                            "Could not retrieve message content for cache invalidation",
+                            layer=self.name,
+                            conversation_id=conversation_id,
+                            message_id=target_message_id,
+                        )
+                else:
+                    logger.warning(
+                        "Could not find message_id for cache invalidation",
+                        layer=self.name,
+                        conversation_id=conversation_id,
+                    )
+            else:
+                logger.warning(
+                    "No workspace_client available for cache invalidation",
+                    layer=self.name,
+                    conversation_id=conversation_id,
+                )
+
+        # Forward feedback to underlying service if not a cache hit
+        # For cache hits, there's no Genie message to provide feedback on
+        if was_cache_hit:
+            logger.info(
+                "Skipping Genie API feedback - response was served from cache",
+                layer=self.name,
+                conversation_id=conversation_id,
+                rating=rating.value if rating else None,
+                cache_invalidated=invalidated,
+            )
+            return
+
+        # Forward to underlying service
+        logger.debug(
+            "Forwarding feedback to underlying service",
+            layer=self.name,
+            conversation_id=conversation_id,
+            rating=rating.value if rating else None,
+            delegating_to=type(self.impl).__name__,
+        )
+        self.impl.send_feedback(
+            conversation_id=conversation_id,
+            rating=rating,
+            message_id=message_id,
+            was_cache_hit=False,  # Already handled, so pass False
+        )