reworked saving/loading agent by just saving the memory documents themselves and adding them on load, rather than serializing the entire memory object and it breaking between systems / wizard prompt tune / more tunings

README.md

@@ -44,7 +44,7 @@ I ***do not*** plan on making this uber-user friendly like [mrq/ai-voice-cloning
 
 ## Caveats
 
-A local LM is quite slow. Things seem to be getting faster as llama.cpp is being developed.
+A local LM is quite slow. Things seem to be getting faster as llama.cpp is being developed. GPU offloading (and the OpenCL PR) seems to bring some very nice hope in just scrapping this in Python and just integrate it entirely in C++.
 
 Even using one that's more instruction-tuned like Vicuna (with a `SYSTEM:\nUSER:\nASSISTANT:` structure of prompts), it's still inconsistent.
 

src/ext/generative_agent.py

@@ -122,14 +122,12 @@ class GenerativeAgent(BaseModel):
         memory = ""
 
         for mem in relevant_memories:
-            if mem in summary or mem in memory:
+            if mem in summary or mem in memory or mem in observation:
                 continue
             memory += f"\n{mem}"
         
         for mem in recent_memories:
-            if mem in summary:
-                continue
-            if mem is observation:
+            if mem in summary or mem in observation:
                 continue
             # erase it, move it to bottom
             if mem in memory:
@@ -169,7 +167,7 @@ class GenerativeAgent(BaseModel):
 
         if self.verbose:
             print(reaction)
-        return reaction
+        return f'{self.name}: {reaction}'
 
     def generate_response(self, observation: str) -> Tuple[bool, str]:
         """React to a given observation."""
@@ -195,7 +193,7 @@ class GenerativeAgent(BaseModel):
         summary = self.chain(prompt).run(stop=get_stop_tokens(), name=self.name, summary=self.summaries[-1] if len(self.summaries) else self.summary, queries=[f"{self.name}'s core characteristics"]).strip()
         if self.verbose:
             print(summary)
-        return summary
+        return f'{self.name} {summary}'
 
     def get_summary(self, force_refresh: bool = False) -> str:
         """Return a descriptive summary of the agent."""

src/ext/prompts.py

@@ -61,8 +61,7 @@ PROMPTS = {
     },
     "generate_reaction": {
         "system": (
-            "[Write one reply. Always stay in character. Maintain a casual tone using beige prose. Be brief. Avoid repeating anything below.]"
-            "\nCurrent Time: {current_time}"
+            #"\nCurrent Time: {current_time}" # commented out, not necessary if I'm not passing time anyways, and I think bigger LLMs would only take advantage of it / llama's prompt caching will get ruined with this changing
             "\n{summary}"
             "\n{memory}"
             "\n{observation}"
@@ -70,7 +69,7 @@ PROMPTS = {
         "user": (
             "{suffix}"
         ),
-        "assistant": ""
+        "assistant": "{name}: "
     },
 
     #
@@ -88,6 +87,7 @@ PROMPT_TUNES = {
     "supercot": "{role}:\n{query}",
     "alpasta": "{role}# {query}",
     "cocktail": "{role}: {query}",
+    "wizard-vicuna": "{role}: {query}",
 }
 PROMPT_ROLES = {
     "vicuna": {
@@ -100,6 +100,11 @@ PROMPT_ROLES = {
         "user": "### Input",
         "assistant": "### Response",
     },
+    "wizard-vicuna": {
+        "system": "### Instruction",
+        "user": "### Input",
+        "assistant": "### Response",
+    },
     "alpasta": {
         "system": "<|system|>",
         "user": "<|user|>",
@@ -173,4 +178,5 @@ def get_prompt( key, tune=LLM_PROMPT_TUNE ):
     output = "\n".join(outputs)
     #if LLM_PROMPT_TUNE == "cocktail":
     output = output.strip()
+    print([output[-1]])
     return output

src/main.py

@@ -78,7 +78,7 @@ def get_summary_proxy( agents ):
 		messages.append(get_summary( agent, force_refresh = True ))
 	return "\n".join(messages)
 
-def run_conversation_proxy( agents, message, limit=4 ):
+def run_conversation_proxy( agents, message, limit=32 ):
 	agents = [ AGENTS[agent] for agent in agents ]
 
 	if len(agents) < 2:
@@ -95,7 +95,6 @@ def run_conversation_proxy( agents, message, limit=4 ):
 		importance_score = agent_observes( agent, [ message ], importance_score=importance_score )[0][0]
 	agents = agents[1:] + [agents[0]]
 
-	dialogue = []
 	while True:
 		for agent in agents:
 			message = agent_reacts( agent, [ message ] )[0]

src/utils.py

@@ -27,28 +27,33 @@ from langchain.vectorstores import FAISS
 LLM_TYPE = os.environ.get('LLM_TYPE', "llamacpp") # options: llamacpp, oai
 LLM_LOCAL_MODEL = os.environ.get('LLM_MODEL', 
 	#"./models/ggml-vicuna-13b-1.1/ggml-vic13b-uncensored-q4_2.bin"
-	"./models/ggml-vicuna-13b-cocktail-v1-q5_0.bin"
+	#"./models/ggml-vicuna-13b-cocktail-v1-q5_0.bin"
+	#"./models/WizardML-Unc-13b-Q5_1.bin"
 	#"./models/llama-13b-supercot-ggml/ggml-model-q4_2.bin"
 	#"./models/llama-33b-supercot-ggml/ggml-model-q4_2.bin"
 	#"./models/gpt4-x-alpasta-30b-ggml-q4_1.bin"
+
+	"./models/Wizard-Vicuna-13B-Uncensored.ggml.q5_1.bin"
+	#"./models/wizardlm-13b-uncensored-ggml-q5_1.bin"
 )
 LLM_CONTEXT = int(os.environ.get('LLM_CONTEXT', '2048'))
 LLM_THREADS = int(os.environ.get('LLM_THREADS', '6'))
+LLM_GPU_LAYERS = int(os.environ.get('LLM_GPU_LAYERS', '99'))
 LLM_TEMPERATURE = float(os.environ.get('LLM_TEMPERATURE', '0.99'))
 EMBEDDING_TYPE = os.environ.get("LLM_EMBEDDING_TYPE", "hf") # options: llamacpp, oai, hf
 
-#LLM_TYPE="oai"
-#os.environ['OPENAI_API_BASE']="https://oai.ecker.tech/proxy/openai"
-#os.environ['OPENAI_API_KEY']=""
-
 # deduce a default given a model path
 if LLM_TYPE=="oai":
 	LLM_PROMPT_TUNE_DEFAULT = "oai"
 else:
 	if "supercot" in LLM_LOCAL_MODEL.lower():
 		LLM_PROMPT_TUNE_DEFAULT = "supercot"
+	if "wizard-vicuna" in LLM_LOCAL_MODEL.lower():
+		LLM_PROMPT_TUNE_DEFAULT = "wizard-vicuna"
 	elif "vicuna" in LLM_LOCAL_MODEL.lower():
 		LLM_PROMPT_TUNE_DEFAULT = "vicuna"
+	elif "wizard" in LLM_LOCAL_MODEL.lower():
+		LLM_PROMPT_TUNE_DEFAULT = "vicuna"
 	elif "alpasta" in LLM_LOCAL_MODEL.lower():
 		LLM_PROMPT_TUNE_DEFAULT = "alpasta"
 	elif "cocktail" in LLM_LOCAL_MODEL.lower():
@@ -72,6 +77,7 @@ if LLM_TYPE=="llamacpp":
 		callback_manager=callback_manager,
 		verbose=True,
 		n_ctx=LLM_CONTEXT,
+		n_gpu_layers=LLM_GPU_LAYERS,
 		temperature=LLM_TEMPERATURE,
 		#n_threads=LLM_THREADS,
 		#use_mlock=True,
@@ -176,8 +182,9 @@ def save_agent( agent ):
 		"status": agent.status,
 		"summary": agent.summary,
 		"summaries": agent.summaries,
-		"memory_retriever": agent.memory.memory_retriever,
+		"memories": [ memory for memory in agent.memory.memory_retriever.memory_stream ]
 	}
+
 	path = f"./agents/{agent.name}.pth"
 	pickle.dump(obj, open(path, 'wb'))
 	print(f"Saved agent:", path)
@@ -186,7 +193,11 @@ def load_agent( name ):
 	path = f"./agents/{name}.pth"
 	obj = pickle.load(open(path, 'rb'))
 	agent = create_agent(**obj)
-	agent.memory.memory_retriever = obj["memory_retriever"]
+	#agent.memory.memory_retriever.memory_stream = obj["memories"]
+	for memory in obj["memories"]:
+		print("Loaded memory:", memory)
+		agent.memory.memory_retriever.add_documents([memory])
+
 	print(f"Loaded agent:", path)
 	return agent