memory_logic.py

# memory_logic.py
import os
import json
import time
from datetime import datetime
import logging
import re
import threading

# Conditionally import heavy dependencies
try:
    from sentence_transformers import SentenceTransformer
    import faiss
    import numpy as np
except ImportError:
    SentenceTransformer, faiss, np = None, None, None
    logging.warning("SentenceTransformers, FAISS, or NumPy not installed. Semantic search will be unavailable.")

try:
    import sqlite3
except ImportError:
    sqlite3 = None
    logging.warning("sqlite3 module not available. SQLite backend will be unavailable.")

try:
    from datasets import load_dataset, Dataset
except ImportError:
    load_dataset, Dataset = None, None
    logging.warning("datasets library not installed. Hugging Face Dataset backend will be unavailable.")


logger = logging.getLogger(__name__)
# Suppress verbose logs from dependencies
for lib_name in ["sentence_transformers", "faiss", "datasets", "huggingface_hub"]:
    if logging.getLogger(lib_name): # Check if logger exists
        logging.getLogger(lib_name).setLevel(logging.WARNING)


# --- Configuration (Read directly from environment variables) ---
STORAGE_BACKEND = os.getenv("STORAGE_BACKEND", "HF_DATASET").upper() #HF_DATASET, RAM, SQLITE
SQLITE_DB_PATH = os.getenv("SQLITE_DB_PATH", "app_data/ai_memory.db") # Changed default path
HF_TOKEN = os.getenv("HF_TOKEN")
HF_MEMORY_DATASET_REPO = os.getenv("HF_MEMORY_DATASET_REPO", "broadfield-dev/ai-brain") # Example
HF_RULES_DATASET_REPO = os.getenv("HF_RULES_DATASET_REPO", "broadfield-dev/ai-rules")     # Example

# --- Globals for RAG within this module ---
_embedder = None
_dimension = 384 # Default, will be set by embedder
_faiss_memory_index = None
_memory_items_list = []  # Stores JSON strings of memory objects for RAM, or loaded from DB/HF
_faiss_rules_index = None
_rules_items_list = []   # Stores rule text strings

_initialized = False
_init_lock = threading.Lock()

# --- Helper: SQLite Connection ---
def _get_sqlite_connection():
    if not sqlite3:
        raise ImportError("sqlite3 module is required for SQLite backend but not found.")
    db_dir = os.path.dirname(SQLITE_DB_PATH)
    if db_dir and not os.path.exists(db_dir):
        os.makedirs(db_dir, exist_ok=True)
    return sqlite3.connect(SQLITE_DB_PATH, timeout=10)

def _init_sqlite_tables():
    if STORAGE_BACKEND != "SQLITE" or not sqlite3:
        return
    try:
        with _get_sqlite_connection() as conn:
            cursor = conn.cursor()
            cursor.execute("""
                CREATE TABLE IF NOT EXISTS memories (
                    id INTEGER PRIMARY KEY AUTOINCREMENT,
                    memory_json TEXT NOT NULL,
                    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
                )
            """)
            cursor.execute("""
                CREATE TABLE IF NOT EXISTS rules (
                    id INTEGER PRIMARY KEY AUTOINCREMENT,
                    rule_text TEXT NOT NULL UNIQUE,
                    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
                )
            """)
            conn.commit()
        logger.info("SQLite tables for memories and rules checked/created.")
    except Exception as e:
        logger.error(f"SQLite table initialization error: {e}", exc_info=True)


def _build_faiss_index(items_list, text_extraction_fn):
    """Builds a FAISS index from a list of items."""
    if not _embedder:
        logger.error("Cannot build FAISS index: Embedder not available.")
        return None
    
    index = faiss.IndexFlatL2(_dimension)
    if not items_list:
        return index

    logger.info(f"Building FAISS index for {len(items_list)} items...")
    texts_to_embed = [text_extraction_fn(item) for item in items_list]
    
    try:
        embeddings = _embedder.encode(texts_to_embed, convert_to_tensor=False, show_progress_bar=False)
        embeddings_np = np.array(embeddings, dtype=np.float32)
        if embeddings_np.ndim == 2 and embeddings_np.shape[0] == len(items_list):
            index.add(embeddings_np)
            logger.info(f"FAISS index built successfully with {index.ntotal} items.")
        else:
            logger.error(f"FAISS build failed: Embeddings shape error. Expected ({len(items_list)}, {_dimension}), Got {getattr(embeddings_np, 'shape', 'N/A')}")
            return faiss.IndexFlatL2(_dimension) # Return empty index on failure
    except Exception as e:
        logger.error(f"Error building FAISS index: {e}", exc_info=True)
        return faiss.IndexFlatL2(_dimension) # Return empty index on failure

    return index

# --- Initialization ---
def initialize_memory_system():
    global _initialized, _embedder, _dimension, _faiss_memory_index, _memory_items_list, _faiss_rules_index, _rules_items_list
    
    with _init_lock:
        if _initialized:
            return

        logger.info(f"Initializing memory system with backend: {STORAGE_BACKEND}")
        init_start_time = time.time()

        if not SentenceTransformer or not faiss or not np:
            logger.error("Core RAG libraries not available. Cannot initialize semantic memory.")
            return
        
        if not _embedder:
            try:
                logger.info("Loading SentenceTransformer model (all-MiniLM-L6-v2)...")
                _embedder = SentenceTransformer('all-MiniLM-L6-v2', cache_folder="./sentence_transformer_cache")
                _dimension = _embedder.get_sentence_embedding_dimension() or 384
            except Exception as e:
                logger.critical(f"FATAL: Error loading SentenceTransformer: {e}", exc_info=True)
                return

        if STORAGE_BACKEND == "SQLITE": _init_sqlite_tables()

        # Load Memories from persistent storage
        temp_memories_json = []
        if STORAGE_BACKEND == "SQLITE":
            try: temp_memories_json = [row[0] for row in _get_sqlite_connection().execute("SELECT memory_json FROM memories")]
            except Exception as e: logger.error(f"Error loading memories from SQLite: {e}")
        elif STORAGE_BACKEND == "HF_DATASET":
            try:
                logger.info(f"Loading memories from HF Dataset: {HF_MEMORY_DATASET_REPO}")
                dataset = load_dataset(HF_MEMORY_DATASET_REPO, token=HF_TOKEN, trust_remote_code=True)
                if "train" in dataset and "memory_json" in dataset["train"].column_names:
                    temp_memories_json = [m for m in dataset["train"]["memory_json"] if isinstance(m, str) and m.strip()]
                    logger.info(f"Loaded {len(temp_memories_json)} valid memories from HF Dataset.")
                else: logger.warning(f"HF Dataset {HF_MEMORY_DATASET_REPO} has no 'train' split or 'memory_json' column.")
            except Exception as e: logger.error(f"Error loading memories from HF Dataset: {e}", exc_info=True)
        
        _memory_items_list = temp_memories_json
        
        # Build Memory FAISS Index
        _faiss_memory_index = _build_faiss_index(
            _memory_items_list, 
            lambda m: f"User: {json.loads(m).get('user_input', '')}\nAI: {json.loads(m).get('bot_response', '')}\nTakeaway: {json.loads(m).get('metrics', {}).get('takeaway', 'N/A')}"
        )

        # Load Rules from persistent storage
        temp_rules_text = []
        if STORAGE_BACKEND == "SQLITE":
            try: temp_rules_text = [row[0] for row in _get_sqlite_connection().execute("SELECT rule_text FROM rules")]
            except Exception as e: logger.error(f"Error loading rules from SQLite: {e}")
        elif STORAGE_BACKEND == "HF_DATASET":
            try:
                logger.info(f"Loading rules from HF Dataset: {HF_RULES_DATASET_REPO}")
                dataset = load_dataset(HF_RULES_DATASET_REPO, token=HF_TOKEN, trust_remote_code=True)
                if "train" in dataset and "rule_text" in dataset["train"].column_names:
                    temp_rules_text = [r for r in dataset["train"]["rule_text"] if isinstance(r, str) and r.strip()]
                    logger.info(f"Loaded {len(temp_rules_text)} valid rules from HF Dataset.")
                else: logger.warning(f"HF Dataset {HF_RULES_DATASET_REPO} has no 'train' split or 'rule_text' column.")
            except Exception as e: logger.error(f"Error loading rules from HF Dataset: {e}", exc_info=True)

        _rules_items_list = sorted(list(set(temp_rules_text)))
        
        # Build Rules FAISS Index
        _faiss_rules_index = _build_faiss_index(_rules_items_list, lambda r: r)

        _initialized = True
        logger.info(f"Memory system initialization complete in {time.time() - init_start_time:.2f}s")


def _verify_and_rebuild_if_needed(index, items_list, text_extraction_fn):
    """Self-healing function to ensure FAISS index is synced with the item list."""
    if not index or index.ntotal != len(items_list):
        logger.warning(
            f"FAISS index mismatch detected (Index: {index.ntotal if index else 'None'}, List: {len(items_list)}). "
            "Rebuilding index from in-memory cache."
        )
        return _build_faiss_index(items_list, text_extraction_fn)
    return index


# --- Memory Operations (Semantic) ---
def add_memory_entry(user_input: str, metrics: dict, bot_response: str) -> tuple[bool, str]:
    global _memory_items_list, _faiss_memory_index
    if not _initialized: initialize_memory_system()
    if not _embedder or not _faiss_memory_index:
        return False, "Memory system not ready for adding entries."

    memory_obj = {"user_input": user_input, "metrics": metrics, "bot_response": bot_response, "timestamp": datetime.utcnow().isoformat()}
    memory_json_str = json.dumps(memory_obj)
    text_to_embed = f"User: {user_input}\nAI: {bot_response}\nTakeaway: {metrics.get('takeaway', 'N/A')}"
    
    try:
        embedding = _embedder.encode([text_to_embed], convert_to_tensor=False)
        embedding_np = np.array(embedding, dtype=np.float32)
        
        _faiss_memory_index.add(embedding_np)
        _memory_items_list.append(memory_json_str)
        
        if STORAGE_BACKEND == "SQLITE":
            with _get_sqlite_connection() as conn:
                conn.execute("INSERT INTO memories (memory_json) VALUES (?)", (memory_json_str,)); conn.commit()
        elif STORAGE_BACKEND == "HF_DATASET":
            Dataset.from_dict({"memory_json": list(_memory_items_list)}).push_to_hub(HF_MEMORY_DATASET_REPO, token=HF_TOKEN, private=True)
        
        logger.info(f"Added memory. Cache size: {len(_memory_items_list)}, FAISS size: {_faiss_memory_index.ntotal}")
        return True, "Memory added successfully."
    except Exception as e:
        logger.error(f"Error adding memory entry: {e}", exc_info=True)
        return False, f"Error adding memory: {e}"

def retrieve_memories_semantic(query: str, k: int = 3) -> list[dict]:
    global _faiss_memory_index
    if not _initialized: initialize_memory_system()
    
    # Self-healing: Verify index is synced with cache, rebuild if not.
    _faiss_memory_index = _verify_and_rebuild_if_needed(
        _faiss_memory_index, _memory_items_list,
        lambda m: f"User: {json.loads(m).get('user_input', '')}\nAI: {json.loads(m).get('bot_response', '')}\nTakeaway: {json.loads(m).get('metrics', {}).get('takeaway', 'N/A')}"
    )

    if not _faiss_memory_index or _faiss_memory_index.ntotal == 0:
        logger.debug("Cannot retrieve memories: index is empty.")
        return []

    try:
        query_embedding = _embedder.encode([query], convert_to_tensor=False)
        query_embedding_np = np.array(query_embedding, dtype=np.float32)
        distances, indices = _faiss_memory_index.search(query_embedding_np, min(k, _faiss_memory_index.ntotal))
        
        results = [json.loads(_memory_items_list[i]) for i in indices[0] if 0 <= i < len(_memory_items_list)]
        logger.info(f"Retrieved {len(results)} memories for query: '{query[:50]}...'")
        return results
    except Exception as e:
        logger.error(f"Error retrieving memories semantically: {e}", exc_info=True)
        return []


# --- Rule (Insight) Operations (Semantic) ---
def add_rule_entry(rule_text: str) -> tuple[bool, str]:
    global _rules_items_list, _faiss_rules_index
    if not _initialized: initialize_memory_system()

    rule_text = rule_text.strip()
    if not rule_text or "duplicate" == rule_text or rule_text in _rules_items_list:
        return False, "duplicate or invalid"
    if not re.match(r"\[(CORE_RULE|RESPONSE_PRINCIPLE|BEHAVIORAL_ADJUSTMENT|GENERAL_LEARNING)\|([\d\.]+?)\]", rule_text, re.I):
        return False, "Invalid rule format."

    try:
        embedding = _embedder.encode([rule_text], convert_to_tensor=False)
        embedding_np = np.array(embedding, dtype=np.float32)
        _faiss_rules_index.add(embedding_np)
        _rules_items_list.append(rule_text)
        _rules_items_list.sort()

        if STORAGE_BACKEND == "SQLITE":
            with _get_sqlite_connection() as conn:
                conn.execute("INSERT OR IGNORE INTO rules (rule_text) VALUES (?)", (rule_text,)); conn.commit()
        elif STORAGE_BACKEND == "HF_DATASET":
            Dataset.from_dict({"rule_text": list(_rules_items_list)}).push_to_hub(HF_RULES_DATASET_REPO, token=HF_TOKEN, private=True)
        return True, "Rule added successfully."
    except Exception as e:
        logger.error(f"Error adding rule entry: {e}", exc_info=True)
        return False, f"Error adding rule: {e}"

def retrieve_rules_semantic(query: str, k: int = 5) -> list[str]:
    global _faiss_rules_index
    if not _initialized: initialize_memory_system()

    _faiss_rules_index = _verify_and_rebuild_if_needed(_faiss_rules_index, _rules_items_list, lambda r: r)

    if not _faiss_rules_index or _faiss_rules_index.ntotal == 0: return []
    try:
        query_embedding = _embedder.encode([query], convert_to_tensor=False)
        query_embedding_np = np.array(query_embedding, dtype=np.float32)
        distances, indices = _faiss_rules_index.search(query_embedding_np, min(k, _faiss_rules_index.ntotal))
        return [_rules_items_list[i] for i in indices[0] if 0 <= i < len(_rules_items_list)]
    except Exception as e:
        logger.error(f"Error retrieving rules semantically: {e}", exc_info=True)
        return []

def remove_rule_entry(rule_text_to_delete: str) -> bool:
    global _rules_items_list, _faiss_rules_index
    if not _initialized: initialize_memory_system()
    rule_text_to_delete = rule_text_to_delete.strip()
    if rule_text_to_delete not in _rules_items_list: return False
    try:
        _rules_items_list.remove(rule_text_to_delete)
        _faiss_rules_index = _build_faiss_index(_rules_items_list, lambda r: r)
        
        if STORAGE_BACKEND == "SQLITE":
            with _get_sqlite_connection() as conn:
                conn.execute("DELETE FROM rules WHERE rule_text = ?", (rule_text_to_delete,)); conn.commit()
        elif STORAGE_BACKEND == "HF_DATASET":
            Dataset.from_dict({"rule_text": list(_rules_items_list)}).push_to_hub(HF_RULES_DATASET_REPO, token=HF_TOKEN, private=True)
        return True
    except Exception as e:
        logger.error(f"Error removing rule entry: {e}", exc_info=True)
        return False

# --- Utility functions to get all data (for UI display, etc.) ---
def get_all_rules_cached() -> list[str]:
    if not _initialized: initialize_memory_system()
    return list(_rules_items_list)

def get_all_memories_cached() -> list[dict]:
    if not _initialized: initialize_memory_system()
    return [json.loads(m) for m in _memory_items_list if m]

def clear_all_memory_data_backend() -> bool:
    global _memory_items_list, _faiss_memory_index
    if not _initialized: initialize_memory_system()
    _memory_items_list = []
    if _faiss_memory_index: _faiss_memory_index.reset()
    try:
        if STORAGE_BACKEND == "SQLITE":
            with _get_sqlite_connection() as conn: conn.execute("DELETE FROM memories"); conn.commit()
        elif STORAGE_BACKEND == "HF_DATASET":
            Dataset.from_dict({"memory_json": []}).push_to_hub(HF_MEMORY_DATASET_REPO, token=HF_TOKEN, private=True)
        logger.info("All memories cleared.")
        return True
    except Exception as e:
        logger.error(f"Error clearing all memory data: {e}"); return False

def clear_all_rules_data_backend() -> bool:
    global _rules_items_list, _faiss_rules_index
    if not _initialized: initialize_memory_system()
    _rules_items_list = []
    if _faiss_rules_index: _faiss_rules_index.reset()
    try:
        if STORAGE_BACKEND == "SQLITE":
            with _get_sqlite_connection() as conn: conn.execute("DELETE FROM rules"); conn.commit()
        elif STORAGE_BACKEND == "HF_DATASET":
            Dataset.from_dict({"rule_text": []}).push_to_hub(HF_RULES_DATASET_REPO, token=HF_TOKEN, private=True)
        logger.info("All rules cleared.")
        return True
    except Exception as e:
        logger.error(f"Error clearing all rules data: {e}"); return False

FAISS_MEMORY_PATH = os.path.join(os.getenv("FAISS_STORAGE_PATH", "app_data/faiss_indices"), "memory_index.faiss")
FAISS_RULES_PATH = os.path.join(os.getenv("FAISS_STORAGE_PATH", "app_data/faiss_indices"), "rules_index.faiss")

def save_faiss_indices_to_disk():
    if not _initialized or not faiss: return
    faiss_dir = os.path.dirname(FAISS_MEMORY_PATH)
    if not os.path.exists(faiss_dir): os.makedirs(faiss_dir, exist_ok=True)
    if _faiss_memory_index and _faiss_memory_index.ntotal > 0:
        faiss.write_index(_faiss_memory_index, FAISS_MEMORY_PATH)
    if _faiss_rules_index and _faiss_rules_index.ntotal > 0:
        faiss.write_index(_faiss_rules_index, FAISS_RULES_PATH)

def load_faiss_indices_from_disk():
    global _faiss_memory_index, _faiss_rules_index
    if not _initialized or not faiss: return
    if os.path.exists(FAISS_MEMORY_PATH):
        _faiss_memory_index = faiss.read_index(FAISS_MEMORY_PATH)
    if os.path.exists(FAISS_RULES_PATH):
        _faiss_rules_index = faiss.read_index(FAISS_RULES_PATH)