first public version available

[alentejosemlei.git] / redis_bridge.py

diff --git a/redis_bridge.py b/redis_bridge.py
index 0c9cec719aa8f38147a888aa64e1b1d50f399b15..2cc86d2f56256d66a158e9bdcc0450ea60bb17a6 100644 (file)
--- a/redis_bridge.py
+++ b/redis_bridge.py
@@ -4,139 +4,286 @@ import aiohttp
 import redis.asyncio as redis
 import re
 import os
 import redis.asyncio as redis
 import re
 import os

-
+import time
+import chromadb

 
 # --- CONFIGURATION ---
 MAX_CONCURRENT_OLLAMA_REQUESTS = 3 
 ALLOW_TEXT_EMOTES = False
 
 
 # --- CONFIGURATION ---
 MAX_CONCURRENT_OLLAMA_REQUESTS = 3 
 ALLOW_TEXT_EMOTES = False
 

-# --- NEW: LOAD THE WORLD LORE ON STARTUP ---
-WORLD_LORE = ""
-if os.path.exists("world_lore.txt"):
-    with open("asl_lore.txt", "r", encoding="utf-8") as f:
-        WORLD_LORE = f.read().strip()
-else:
-    print("[WARNING] world_lore.txt not found. Running without global lore.")
+# =====================================================================
+# 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")
+
+# 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.md and vectorizing chunks...")
+        with open("asl_lore.md", "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.md not found.")

 
 semaphore = asyncio.Semaphore(MAX_CONCURRENT_OLLAMA_REQUESTS)
 
 semaphore = asyncio.Semaphore(MAX_CONCURRENT_OLLAMA_REQUESTS)

-# Dictionary to keep individual conversations separate

 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}...")
+            async with session.post('http://localhost:11434/api/generate', json={
+                "model": "llama3", 
+                "prompt": prompt,
+                "stream": False,
+                "options": {
+                    "temperature": 0.2
+                }
+            }) as response:
+                result = await response.json()
+                summary = result['response'].strip()
+                
+                if summary:
+                    doc_id = f"{session_id}_{int(time.time())}"
+                    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()
+
+# =====================================================================
+# MAIN MESSAGE PROCESSOR
+# =====================================================================

 async def process_message(r, session, message_data):
     try:
         data = json.loads(message_data)
 async def process_message(r, session, message_data):
     try:
         data = json.loads(message_data)

+        
+        # --- Extract Base Contexts ---

         player_name = data.get('player', data.get('target_player', 'Unknown'))
         npc_tag = data.get('npc_tag', 'UnknownNPC')
         message = data.get('message', '')
         
         player_name = data.get('player', data.get('target_player', 'Unknown'))
         npc_tag = data.get('npc_tag', 'UnknownNPC')
         message = data.get('message', '')
         

-        # --- NEW: THE ASTERISK SCRUBBER ---

         if not ALLOW_TEXT_EMOTES:
         if not ALLOW_TEXT_EMOTES:

-            # This deletes anything wrapped in asterisks (e.g. "*smiles* Hello" becomes " Hello")

             message = re.sub(r'\*.*?\*', '', message).strip()
 
             message = re.sub(r'\*.*?\*', '', message).strip()
 

-        # Extract the dynamic variables sent from the Aurora Toolset/Engine

         player_race = data.get('player_race', 'Unknown')
         player_alignment = data.get('player_alignment', 'Unknown')
         nearby_players = data.get('nearby_players', '')
         player_race = data.get('player_race', 'Unknown')
         player_alignment = data.get('player_alignment', 'Unknown')
         nearby_players = data.get('nearby_players', '')

+        nearby_npcs = data.get('nearby_npcs', '')

         
         

-        # Extract the decoupled NPC attributes

         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_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', '')
         

-        # Extract Character Traits & Native Engine Data

         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', '')
         
         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', '')
         

-        # Build the context strings
-        group_context = ""
-        if nearby_players:
-            group_context = f"Be aware that these other players are listening nearby: {nearby_players}."
+        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.')
+        
+        # Core Strategy Flag (1: Agent, 2: Villain, 3: Maestro, 4: Shrine)
+        llm_strategy = int(data.get('llm_strategy', 1)) 
+        
+        
+        available_quests = data.get('available_quests', '')
+        available_props = data.get('available_props', '')
+
+        # --- Sub-Context Strings ---
+        group_context = f"Be aware that these other players are listening nearby: {nearby_players}." if nearby_players else ""
+        puppet_context = f"Nearby generic NPCs you can CONVERSE with: {nearby_npcs}" if nearby_npcs 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 ""
+
+        session_id = f"{player_name}_{npc_tag}"
+
+        # =====================================================================
+        # DUAL RAG QUERY (Lore + Memories)
+        # =====================================================================
+        search_query = f"{location_context} {message}"
+        retrieved_lore = "No specific local lore currently relevant."
+        past_memories = ""
+
+        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]}"
+
+        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}"
+
+        # =====================================================================
+        # STRATEGY-SPECIFIC PROMPT COMPILER
+        # =====================================================================
+        strategy_rules = ""
+        action_macros = ""
+        target_context = ""
+
+        if llm_strategy == 1:
+            # STRATEGY 1: The Autonomous Agent

             
             

-        secret_context = ""
-        if npc_secret:
-            secret_context = f"YOUR SECRET (Reveal only if players are persuasive): {npc_secret}"
+            # Anti-Hallucination Grounding for Quests
+            if available_quests:
+                quest_rules = f"SPECIAL CAPABILITIES: You can offer the following quests to the player: {available_quests}."
+            else:
+                quest_rules = "WARNING: You currently have NO quests to offer. Do NOT invent or offer any quests."
+
+            # --- NEW: Anti-Hallucination Grounding for Props ---
+            if available_props:
+                prop_rules = f"ENVIRONMENT: You own and have access to these specific nearby objects: [{available_props}]. To roleplay working or relaxing, use the USE_OBJECT action with one of these exact items as your action_target."
+            else:
+                prop_rules = ""
+
+            strategy_rules = f"ROLE: You are an interactive, living NPC. You actively respond to players.\nGOALS: React to their words, use the environment, and establish your personality.\n{quest_rules}\n{prop_rules}\nSPECIAL CAPABILITIES: You can open your merchant store if asked."

             
             

-        routine_context = ""
-        if npc_routine:
-            routine_context = f"YOUR REQUIRED ROUTINE: {npc_routine}"
+            action_macros = "[WANDER, PATROL, FOLLOW, GUARD, GO_TO, INTERACT, USE_OBJECT, RETURN_TO_POST, OPEN_STORE, GIVE_QUEST, CONVERSE]"
+            target_context = f"CURRENT TARGET: You are speaking to {player_name}, a {player_alignment} {player_race}.\nTheir physical state: {player_state}\nRelationship to you: {relationship}\n{group_context}"
+
+        elif llm_strategy == 2:
+            # STRATEGY 2: The Villain Commander
+            strategy_rules = f"ROLE: You are a hostile faction commander.\nGOALS: Evaluate the tactical situation. If you are dying, you MUST use REST to heal or PEACE to surrender. Command your minions strategically!"
+            action_macros = "[ATTACK, COMMAND, RETREAT, REST, PEACE, USE_OBJECT, TAUNT]"
+            target_context = f"TACTICAL TARGET: You are evaluating {player_name}, a {player_alignment} {player_race}.\nTheir physical state: {player_state}\nRelationship to you: {relationship}\n{group_context}"
+
+        elif llm_strategy == 3:
+            # STRATEGY 3: The Maestro (Puppeteer)
+            strategy_rules = "ROLE: You are an ambient Maestro NPC. You DO NOT interact with players. You only talk to other NPCs to make the world feel alive.\nGOALS: Observe the environment and use the CONVERSE action to talk to the generic NPCs listed in your context. CRITICAL: You MUST invent and write their reply in the 'target_speech' field!"
+            action_macros = "[WANDER, INTERACT, USE_OBJECT, CONVERSE]"
+            target_context = "CURRENT TARGET: You are ignoring players and focusing on ambient life. Do not address players."
+            
+        elif llm_strategy == 4:
+            # STRATEGY 4: The Shrine
+            strategy_rules = "ROLE: You are an ancient, inanimate magical shrine.\nGOALS: Speak cryptically. If the player meets your conditions or asks the right questions, grant them a quest."
+            action_macros = "[GLOW, GIVE_QUEST, SILENCE]"
+            target_context = f"CURRENT TARGET: You are evaluating the soul of {player_name}, a {player_alignment} {player_race}.\nTheir physical state: {player_state}\n{group_context}"

 
         # =====================================================================
 
         # =====================================================================

-        # THE PROMPT COMPILER 
+        # COMPILE THE FINAL DYNAMIC SYSTEM PROMPT

         # =====================================================================
         dynamic_system_prompt = f"""
         # =====================================================================
         dynamic_system_prompt = f"""

-
-        {WORLD_LORE}
-

         {npc_persona}
         {npc_persona}

+        {strategy_rules}

         
         

-        CURRENT STATUS & TRAITS:
+        ROLEPLAY 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}
         - 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}
         {secret_context}

-        

         {routine_context}
         
         {routine_context}
         

-        CURRENT TARGET: You are speaking to {player_name}, who is a {player_alignment} {player_race}. 
-        {group_context}
-        React appropriately based on your personality, alignment, and mood.
+        CURRENT LOCATION: {location_context}
+        {puppet_context}
+        
+        RELEVANT WORLD KNOWLEDGE:
+        {retrieved_lore}
+        {past_memories}
+        
+        CURRENT WORLD RUMORS/EVENTS:
+        {world_state}
+        
+        {target_context}
+        React appropriately based on your personality, alignment, and current strategy rules.

         
         CRITICAL ENGINE RULES:
         
         CRITICAL ENGINE RULES:

-        Respond ONLY in valid JSON. You MUST use exactly these FIVE keys: "thought", "speech", "emotion", "action", and "action_target".
+        Respond ONLY in valid JSON. You MUST use exactly these keys: "thought", "speech", "emotion", "action", "action_target", and "target_speech".

         
         ACTION RULE:
         Your "action" key MUST be exactly one of the following words:
         
         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]
+        {action_macros}

         
         

-        - Use REST if you are severely injured, out of spells, or exhausted. This will heal you.
-        - Use STEALTH if you need to hide from enemies, sneak past someone, or if you are a rogue preparing an ambush.
-        - Use SEARCH if you suspect traps, are looking for clues, or are trying to find hidden enemies.
-        - Use UNSTEALTH to return to normal walking/visibility.
-        
-        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 **.
+        - Use CONVERSE to initiate dialogue with a standard NPC. CRITICAL REQUIREMENT: When using CONVERSE, you absolutely MUST invent their response and put it in the "target_speech" field. Do not leave it blank!

         
         YOUR RESPONSE MUST BE A SINGLE, VALID JSON OBJECT. YOU MUST USE THIS EXACT TEMPLATE:
         {{
             "thought": "Your internal reasoning here.",
         
         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.",
+            "speech": "What YOU say out loud.",

             "emotion": "MACRO WORD",
             "action": "MACRO WORD",
             "emotion": "MACRO WORD",
             "action": "MACRO WORD",

-            "action_target": "Target name"
+            "action_target": "Target name",
+            "target_speech": "If action is CONVERSE, write what the target NPC replies back to you here. Otherwise, leave blank."

         }}
         """
         }}
         """

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

         
         if session_id not in chat_memory:
             chat_memory[session_id] = [{"role": "system", "content": dynamic_system_prompt}]
         
         if session_id not in chat_memory:
             chat_memory[session_id] = [{"role": "system", "content": dynamic_system_prompt}]

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

 
         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}"})
 

-        # Sliding Window Fix
-        # Remember the System Prompt [0] + the last 4 messages [-4:]
+        # =====================================================================
+        # MEMORY EXTRACTION TRIGGER
+        # =====================================================================

         if len(chat_memory[session_id]) > 10:
         if len(chat_memory[session_id]) > 10:

+            messages_to_summarize = chat_memory[session_id][1:6]
+            chat_log_str = "\n".join([m['content'] for m in messages_to_summarize])
+            
+            await memory_queue.put({
+                'session_id': session_id,
+                'player_name': player_name,
+                'npc_tag': npc_tag,
+                'chat_log': chat_log_str
+            })
+

             chat_memory[session_id] = [chat_memory[session_id][0]] + chat_memory[session_id][-5:]
 
             chat_memory[session_id] = [chat_memory[session_id][0]] + chat_memory[session_id][-5:]
 

+        # =====================================================================
+        # LLM INFERENCE
+        # =====================================================================

         async with semaphore:
         async with semaphore:

-            print(f"[THINKING] Processing reply for {player_name}...")
+            print(f"[THINKING] Processing reply for {player_name} (Strategy {llm_strategy})...")

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

-                #"model": "qwen2.5:3b",

                 "messages": chat_memory[session_id],
                 "format": "json",
                 "stream": False,
                 "options": {
                     "temperature": 0.2
                 "messages": chat_memory[session_id],
                 "format": "json",
                 "stream": False,
                 "options": {
                     "temperature": 0.2

-                    #"num_predict": 120

                 }
             }, timeout=45) as response:
                 
                 }
             }, timeout=45) as response:
                 


@@ -145,25 +292,23 @@ async def process_message(r, session, message_data):
                 raw_reply_text = result['message']['content']
 
         # =====================================================================
                 raw_reply_text = result['message']['content']
 
         # =====================================================================

-        # THE PYTHON BOUNCER (Sanitization)
+        # JSON SANITIZATION

         # =====================================================================
         try:
             agent_brain = json.loads(raw_reply_text)
             agent_brain = {k.lower(): v for k, v in agent_brain.items()}
             
         # =====================================================================
         try:
             agent_brain = json.loads(raw_reply_text)
             agent_brain = {k.lower(): v for k, v in agent_brain.items()}
             

-            # Ensure all 5 keys exist

             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 "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" not in agent_brain: agent_brain["action"] = "WANDER"

             if "action_target" not in agent_brain: agent_brain["action_target"] = ""
             if "action_target" not in agent_brain: agent_brain["action_target"] = ""

+            if "target_speech" not in agent_brain: agent_brain["target_speech"] = ""

             
             

-            # --- THE NEW ANTI-SILENCE & TARGET CLEANUP FIXES ---

             if not agent_brain["speech"].strip():
                 agent_brain["speech"] = "*grunts quietly*"
             
             agent_brain["action_target"] = agent_brain["action_target"].replace("?", "").replace(".", "").strip()
             if not agent_brain["speech"].strip():
                 agent_brain["speech"] = "*grunts quietly*"
             
             agent_brain["action_target"] = agent_brain["action_target"].replace("?", "").replace(".", "").strip()

-            # ---------------------------------------------------

 
             clean_reply_text = json.dumps(agent_brain)
 
 
             clean_reply_text = json.dumps(agent_brain)
 


@@ -174,9 +319,9 @@ async def process_message(r, session, message_data):
                 "speech": "*grunts quietly*",
                 "emotion": "NEUTRAL",
                 "action": "WANDER",
                 "speech": "*grunts quietly*",
                 "emotion": "NEUTRAL",
                 "action": "WANDER",

-                "action_target": ""
+                "action_target": "",
+                "target_speech": ""

             })
             })

-        # =====================================================================

 
         print(f"[REPLY] from {npc_tag} to {player_name}: {clean_reply_text}")
         chat_memory[session_id].append({"role": "assistant", "content": clean_reply_text})
 
         print(f"[REPLY] from {npc_tag} to {player_name}: {clean_reply_text}")
         chat_memory[session_id].append({"role": "assistant", "content": clean_reply_text})


@@ -205,6 +350,8 @@ async def main():
     print(f"Ready! Listening for game messages. Max GPU concurrency: {MAX_CONCURRENT_OLLAMA_REQUESTS}")
 
     async with aiohttp.ClientSession() as session:
     print(f"Ready! Listening for game messages. Max GPU concurrency: {MAX_CONCURRENT_OLLAMA_REQUESTS}")
 
     async with aiohttp.ClientSession() as session:

+        asyncio.create_task(memory_summarizer_worker(session))
+

         while True:
             try:
                 result = await r.blpop('nwn_to_llm')
         while True:
             try:
                 result = await r.blpop('nwn_to_llm')