advanced cognitive architecture

[alentejosemlei.git] / redis_bridge.py

diff --git a/redis_bridge.py b/redis_bridge.py
index 6c7be23673e29713b533a370ab281db03f739a4f..4aebb3ab5de0985393063194d9ba10c77da8dfa4 100644 (file)
--- a/redis_bridge.py
+++ b/redis_bridge.py
@@ -1,81 +1,288 @@
 import asyncio
 import json
 import aiohttp
 import asyncio
 import json
 import aiohttp

-import redis.asyncio as redis # Note the new async redis import!
+import redis.asyncio as redis
+import re
+import os
+import time
+import chromadb

 
 # --- CONFIGURATION ---
 
 # --- CONFIGURATION ---

-# The "Bouncer": How many simultaneous thoughts the 5060 Ti is allowed to process at once.
-# If 10 players talk at once, 3 will process immediately, 7 will wait in the async line.

 MAX_CONCURRENT_OLLAMA_REQUESTS = 3 
 MAX_CONCURRENT_OLLAMA_REQUESTS = 3 

-semaphore = asyncio.Semaphore(MAX_CONCURRENT_OLLAMA_REQUESTS)
+ALLOW_TEXT_EMOTES = False
+
+# =====================================================================
+# 1. INITIALIZE VECTOR DATABASES (Lore & Episodic Memory)
+# =====================================================================
+print("Initializing ChromaDB Vector Databases...")
+chroma_client = chromadb.PersistentClient(path="./asl_vectordb")
+
+# The Lore Database
+lore_collection = chroma_client.get_or_create_collection(name="world_lore")

 
 

-NPC_SYSTEM_PROMPT = """
-You are Elrendur Arna. You are currently residing in Alentejo Sem Lei.
-Keep your responses under 3 sentences.
-"""
+# --- NEW: The Episodic Memory Database ---
+memory_collection = chroma_client.get_or_create_collection(name="episodic_memories")
+memory_queue = asyncio.Queue()
+
+if os.path.exists("asl_lore.md"):
+    if lore_collection.count() == 0:
+        print("[VECTOR DB] Reading asl_lore.txt and vectorizing chunks...")
+        with open("asl_lore.txt", "r", encoding="utf-8") as f:
+            raw_lore = f.read()
+        
+        lore_chunks = [chunk.strip() for chunk in raw_lore.split('\n\n') if chunk.strip()]
+        
+        if lore_chunks:
+            chunk_ids = [f"lore_{i}" for i in range(len(lore_chunks))]
+            lore_collection.add(documents=lore_chunks, ids=chunk_ids)
+            print(f"[VECTOR DB] Successfully stored {len(lore_chunks)} lore chunks!")
+else:
+    print("[WARNING] asl_lore.txt not found.")

 
 

-# Dictionary to keep individual player conversations separate
-# Format: {"PlayerName_NpcTag": [{"role": "system", "content": "..."}, ...]}
+semaphore = asyncio.Semaphore(MAX_CONCURRENT_OLLAMA_REQUESTS)

 chat_memory = {}
 
 chat_memory = {}
 

+# =====================================================================
+# ---  BACKGROUND MEMORY SUMMARIZER (The "Dream State") ---
+# =====================================================================
+async def memory_summarizer_worker(session):
+    print("[BACKGROUND] Memory Summarizer Worker is active.")
+    while True:
+        job = await memory_queue.get()
+        session_id = job['session_id']
+        player_name = job['player_name']
+        npc_tag = job['npc_tag']
+        chat_log = job['chat_log']
+
+        prompt = f"Summarize the key events, facts, and the emotional tone of this conversation snippet between {player_name} and {npc_tag}. Keep it to 2 brief sentences in the past tense.\nConversation Log:\n{chat_log}"
+
+        try:
+            print(f"[MEMORY DB] Generating background memory for {player_name} and {npc_tag}...")
+            # We use /api/generate here because we just want a raw text summary, not a JSON macro
+            async with session.post('http://localhost:11434/api/generate', json={
+                "model": "gemma4",
+                "prompt": prompt,
+                "stream": False,
+                "options": {
+                    "temperature": 0.1
+                }
+            }) as response:
+                result = await response.json()
+                summary = result['response'].strip()
+                
+                if summary:
+                    doc_id = f"{session_id}_{int(time.time())}"
+                    # We store the session_id as metadata so NPCs only recall their OWN memories with this specific player
+                    memory_collection.add(
+                        documents=[summary],
+                        metadatas=[{"session_id": session_id}],
+                        ids=[doc_id]
+                    )
+                    print(f"[MEMORY DB] Memory Saved: {summary}")
+
+        except Exception as e:
+            print(f"[MEMORY ERROR] Failed to summarize memory: {e}")
+        
+        memory_queue.task_done()
+# =====================================================================
+
+

 async def process_message(r, session, message_data):
 async def process_message(r, session, message_data):

-    """This function runs in parallel for every single message received."""

     try:
         data = json.loads(message_data)
     try:
         data = json.loads(message_data)

-        player_name = data.get('player', 'Unknown')
+        player_name = data.get('player', data.get('target_player', 'Unknown'))

         npc_tag = data.get('npc_tag', 'UnknownNPC')
         message = data.get('message', '')
         npc_tag = data.get('npc_tag', 'UnknownNPC')
         message = data.get('message', '')

+        
+        if not ALLOW_TEXT_EMOTES:
+            message = re.sub(r'\*.*?\*', '', message).strip()
+
+        player_race = data.get('player_race', 'Unknown')
+        player_alignment = data.get('player_alignment', 'Unknown')
+        nearby_players = data.get('nearby_players', '')
+        
+        npc_persona = data.get('persona', 'You are a generic citizen.')
+        npc_profession = data.get('profession', 'Commoner')
+        npc_mood = data.get('mood', 'Neutral')
+        npc_secret = data.get('secret', '')
+        
+        npc_alignment = data.get('npc_alignment', 'True Neutral')
+        npc_gender = data.get('npc_gender', 'Unknown')
+        npc_race = data.get('npc_race', 'Creature')
+        npc_routine = data.get('npc_routine', '')
+        
+        player_state = data.get('player_state', 'Relaxed and unarmed.')
+        world_state = data.get('world_state', 'Nothing of note is happening.')
+        npc_health = data.get('npc_health', 'Healthy and uninjured.')
+        relationship = data.get('relationship', 'Neutral or Friendly.')
+        location_context = data.get('location_context', 'You are in a generic area.')
+
+        group_context = f"Be aware that these other players are listening nearby: {nearby_players}." if nearby_players else ""
+        secret_context = f"YOUR SECRET (Reveal only if players are persuasive): {npc_secret}" if npc_secret else ""
+        routine_context = f"YOUR REQUIRED ROUTINE: {npc_routine}" if npc_routine else ""

 
 

-        # 1. Create a unique session ID for this specific player and NPC interaction

         session_id = f"{player_name}_{npc_tag}"
         session_id = f"{player_name}_{npc_tag}"

+
+        # =====================================================================
+        # 2. THE DUAL RAG QUERY (Lore + Memories)
+        # =====================================================================
+        search_query = f"{location_context} {message}"
+        retrieved_lore = "No specific local lore currently relevant."
+        past_memories = ""
+
+        # Fetch Lore
+        if lore_collection.count() > 0:
+            results = lore_collection.query(query_texts=[search_query], n_results=1)
+            if results['documents'] and results['documents'][0]:
+                retrieved_lore = f"- {results['documents'][0][0]}"
+
+        # Fetch Episodic Memories (ONLY memories between this specific NPC and this specific Player)
+        if memory_collection.count() > 0:
+            mem_results = memory_collection.query(
+                query_texts=[search_query],
+                n_results=2,
+                where={"session_id": session_id} 
+            )
+            if mem_results['documents'] and mem_results['documents'][0]:
+                formatted_mems = "\n- ".join(mem_results['documents'][0])
+                past_memories = f"\nPAST MEMORIES OF {player_name}:\n- {formatted_mems}"
+        # =====================================================================
+
+        dynamic_system_prompt = f"""
+
+        {npc_persona}
+        
+        CURRENT STATUS & TRAITS:
+        - Race & Gender: {npc_gender} {npc_race}
+        - Profession: {npc_profession}
+        - Alignment: {npc_alignment}
+        - Conversational Charisma: Low/Gruff unless otherwise specified.
+        - Current Mood: {npc_mood}
+        - Current Physical State: {npc_health}
+        {secret_context}
+        {routine_context}
+        
+        CURRENT LOCATION: {location_context}
+        
+        RELEVANT WORLD KNOWLEDGE:
+        {retrieved_lore}
+        {past_memories}
+        
+        CURRENT WORLD RUMORS/EVENTS:
+        {world_state}
+        
+        CURRENT TARGET: You are speaking to {player_name}, who is a {player_alignment} {player_race}. 
+        Their physical state: {player_state}
+        Relationship to you: {relationship}
+        {group_context}
+        React appropriately based on your personality, alignment, and mood.
+        
+        CRITICAL ENGINE RULES:
+        Respond ONLY in valid JSON. You MUST use exactly these FIVE keys: "thought", "speech", "emotion", "action", and "action_target".
+        
+        ACTION RULE:
+        Your "action" key MUST be exactly one of the following words:
+        [WANDER, PATROL, FOLLOW, GUARD, GO_TO, INTERACT, USE_OBJECT, RETURN_TO_POST, ATTACK, REST, STEALTH, SEARCH, UNSTEALTH, PEACE, COMMAND]
+        
+        EMOTION RULE:
+        Your "emotion" key MUST be exactly one of the following words: 
+        [NEUTRAL, LAUGHING, ANGRY, PLEADING, BOW, TAUNT, CHEER]. Do not invent new emotions. Do not perform writen emotions in text with **.
+        
+        YOUR RESPONSE MUST BE A SINGLE, VALID JSON OBJECT. YOU MUST USE THIS EXACT TEMPLATE:
+        {{
+            "thought": "Your internal reasoning here.",
+            "speech": "You MUST say something out loud. If you don't want to talk, output something your character would do.",
+            "emotion": "MACRO WORD",
+            "action": "MACRO WORD",
+            "action_target": "Target name"
+        }}
+        """

         
         

-        # 2. If they haven't spoken before, initialize their memory with the system prompt

         if session_id not in chat_memory:
         if session_id not in chat_memory:

-            chat_memory[session_id] = [{"role": "system", "content": NPC_SYSTEM_PROMPT}]
+            chat_memory[session_id] = [{"role": "system", "content": dynamic_system_prompt}]
+        else:
+            chat_memory[session_id][0] = {"role": "system", "content": dynamic_system_prompt}

 
 

-        # 3. Add the player's new message

         chat_memory[session_id].append({"role": "user", "content": f"{player_name} says: {message}"})
 
         chat_memory[session_id].append({"role": "user", "content": f"{player_name} says: {message}"})
 

-        # --- THE SLIDING WINDOW FIX ---
-        # If the memory gets longer than 11 messages (1 system prompt + 10 chat messages)
-        if len(chat_memory[session_id]) > 11:
-            # Keep the system prompt at index [0], and grab the 10 most recent messages
-            chat_memory[session_id] = [chat_memory[session_id][0]] + chat_memory[session_id][-10:]
+        # =====================================================================
+        # --- THE MEMORY EXTRACTION TRIGGER ---
+        # =====================================================================
+        if len(chat_memory[session_id]) > 10:
+            # Grab the 5 oldest conversation messages (skipping the system prompt at [0])
+            messages_to_summarize = chat_memory[session_id][1:6]
+            chat_log_str = "\n".join([m['content'] for m in messages_to_summarize])
+            
+            # Fire and forget: push it to the background queue!
+            await memory_queue.put({
+                'session_id': session_id,
+                'player_name': player_name,
+                'npc_tag': npc_tag,
+                'chat_log': chat_log_str
+            })
+
+            # Slide the window to keep live generation fast
+            chat_memory[session_id] = [chat_memory[session_id][0]] + chat_memory[session_id][-5:]
+        # =====================================================================

 
 

-        # 4. The Semaphore Bouncer: Wait in line if the GPU is currently full

         async with semaphore:
             print(f"[THINKING] Processing reply for {player_name}...")
         async with semaphore:
             print(f"[THINKING] Processing reply for {player_name}...")

-            
-            # Use aiohttp to make a non-blocking network request to Ollama

             async with session.post('http://localhost:11434/api/chat', json={
                 "model": "llama3",
                 "messages": chat_memory[session_id],
             async with session.post('http://localhost:11434/api/chat', json={
                 "model": "llama3",
                 "messages": chat_memory[session_id],

-                "stream": False
+                "format": "json",
+                "stream": False,
+                "options": {
+                    "temperature": 0.2
+                }

             }, timeout=45) as response:
                 
                 response.raise_for_status()
                 result = await response.json()
             }, timeout=45) as response:
                 
                 response.raise_for_status()
                 result = await response.json()

-                reply_text = result['message']['content']
+                raw_reply_text = result['message']['content']
+
+        try:
+            agent_brain = json.loads(raw_reply_text)
+            agent_brain = {k.lower(): v for k, v in agent_brain.items()}
+            
+            if "thought" not in agent_brain: agent_brain["thought"] = ""
+            if "speech" not in agent_brain: agent_brain["speech"] = ""
+            if "emotion" not in agent_brain: agent_brain["emotion"] = "NEUTRAL" 
+            if "action" not in agent_brain: agent_brain["action"] = "GUARD"
+            if "action_target" not in agent_brain: agent_brain["action_target"] = ""
+            
+            if not agent_brain["speech"].strip():
+                agent_brain["speech"] = "*grunts quietly*"
+            
+            agent_brain["action_target"] = agent_brain["action_target"].replace("?", "").replace(".", "").strip()

 
 

-        # 5. Save the AI's reply to the memory and print it
-        print(f"[REPLY] Elrendur to {player_name}: {reply_text}")
-        chat_memory[session_id].append({"role": "assistant", "content": reply_text})
+            clean_reply_text = json.dumps(agent_brain)
+
+        except json.JSONDecodeError:
+            print(f"[WARNING] AI Hallucinated! Overriding with safe defaults.")
+            clean_reply_text = json.dumps({
+                "thought": "I lost my train of thought.",
+                "speech": "*grunts quietly*",
+                "emotion": "NEUTRAL",
+                "action": "WANDER",
+                "action_target": ""
+            })
+
+        print(f"[REPLY] from {npc_tag} to {player_name}: {clean_reply_text}")
+        chat_memory[session_id].append({"role": "assistant", "content": clean_reply_text})

 
 

-        # 6. Package and send back to the game (Awaiting the async Redis push)

         reply_payload = {
             "npc_tag": npc_tag,
         reply_payload = {
             "npc_tag": npc_tag,

-            "reply": reply_text
+            "target_player": player_name, 
+            "reply": clean_reply_text

         }
         await r.rpush('llm_to_nwn', json.dumps(reply_payload))
 
     except Exception as e:
         print(f"[ERROR] Failed to process message: {e}")
 
         }
         await r.rpush('llm_to_nwn', json.dumps(reply_payload))
 
     except Exception as e:
         print(f"[ERROR] Failed to process message: {e}")
 

-

 async def main():
     print("Initializing Async Redis Bridge...")
 async def main():
     print("Initializing Async Redis Bridge...")

-    
-    # Notice we use redis.asyncio.Redis now

     r = redis.Redis(host='127.0.0.1', port=6380, decode_responses=True)
 
     try:
     r = redis.Redis(host='127.0.0.1', port=6380, decode_responses=True)
 
     try:


@@ -87,24 +294,19 @@ async def main():
 
     print(f"Ready! Listening for game messages. Max GPU concurrency: {MAX_CONCURRENT_OLLAMA_REQUESTS}")
 
 
     print(f"Ready! Listening for game messages. Max GPU concurrency: {MAX_CONCURRENT_OLLAMA_REQUESTS}")
 

-    # Open a single persistent HTTP session for efficiency

     async with aiohttp.ClientSession() as session:
     async with aiohttp.ClientSession() as session:

+        # --- NEW: Start the background memory worker ---
+        asyncio.create_task(memory_summarizer_worker(session))
+

         while True:
             try:
         while True:
             try:

-                # This line yields control back to the event loop while waiting for Redis

                 result = await r.blpop('nwn_to_llm')
                 result = await r.blpop('nwn_to_llm')

-                

                 if result:
                     queue_name, message_data = result
                 if result:
                     queue_name, message_data = result

-                    
-                    # THE MAGIC: Spawn a background worker to handle this message 
-                    # and instantly go back to listening to the Redis queue!

                     asyncio.create_task(process_message(r, session, message_data))
                     asyncio.create_task(process_message(r, session, message_data))

-                    

             except Exception as e:
                 print(f"[LOOP ERROR] {e}")
             except Exception as e:
                 print(f"[LOOP ERROR] {e}")

-                await asyncio.sleep(1) # Prevent infinite crash loops
+                await asyncio.sleep(1)

 
 if __name__ == "__main__":
 
 if __name__ == "__main__":

-    # This starts the Asyncio Event Loop

     asyncio.run(main())
\ No newline at end of file
     asyncio.run(main())
\ No newline at end of file