Update Documentation

README.md

@@ -12,7 +12,7 @@ The original version was also created in only 7 days, so it is not exactly very
 - Audio File playback (for pre-generated songs/covers created with something like [RVC](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)
 - Vtube Studio Plugin & Model/Prop control
 - Flexible LLM - Load any model into text-generation-webui (tested) or use any openai-compatible endpoint (not tested).
-- Memory - Long-term (persists across restarts) memories can be manually added, but they will also be 
+- Memory/RAG - Long-term (persists across restarts) memories can be manually added, but they will also be 
 automatically generated as the AI talks. (See memories/readme.md for details)
 
 ## Architecture
@@ -112,20 +112,23 @@ documentation [here](https://pytwitchapi.dev/en/stable/index.html#user-authentic
 
 ### This Project
 
-A virtual environment of some sort is recommended (Python 3.11); this project was developed with venv.
+A virtual environment of some sort is recommended (Python 3.11 required); this project was developed with venv.
 
-Install requirements.txt (This is just a pip freeze, so if you're not on windows watch out)
+Install the CUDA 11.8 version of pytorch 2.2.2 first.
 
-DeepSpeed (For TTS) will probably need to be installed separately, I was using instructions
-from [AllTalkTTS](https://github.com/erew123/alltalk_tts?#-deepspeed-installation-options) , and using their 
+Then install requirements.txt (This is just a pip freeze, so if you're not on Windows watch out)
+
+Finally, DeepSpeed (For TTS) will need to be installed separately. I was using instructions
+from [AllTalkTTS](https://github.com/erew123/alltalk_tts?#-deepspeed-installation-options), and using their 
 [provided wheels](https://github.com/erew123/alltalk_tts/releases/tag/DeepSpeed-14.0).
 
-Create an .env file using .env.example as reference. You need your Twitch app id and secret.
+Create an .env file using .env.example as reference. You need your Twitch app id and secret, along with your
+Hugginface token if you use a gated model (like Llama 3).
 
 Place a voice reference wav file in the voices directory. It should be 5~30 seconds long. For details see the RealtimeTTS
 repository.
 
-Find your desired microphone and speaker device numbers by running utils/listAudioDevices.py and note its number. 
+Find your desired microphone and speaker device numbers by running utils/listAudioDevices.py and note its numbers. 
 
 Configure constants.py.
 

constants.py

@@ -1,10 +1,13 @@
 # This file holds various constants used in the program
+# Variables marked with #UNIQUE# will be unique to your setup and NEED to be changed or the program will not work correctly.
 
 # CORE SECTION: All constants in this section are necessary
 
+# Microphone/Speaker device indices
 # Use utils/listAudioDevices.py to find the correct device ID
+#UNIQUE#
 INPUT_DEVICE_INDEX = 1
-OUTPUT_DEVICE_INDEX = 12
+OUTPUT_DEVICE_INDEX = 7
 
 # How many seconds to wait before prompting AI
 PATIENCE = 60
@@ -17,9 +20,11 @@
 TWITCH_MAX_MESSAGE_LENGTH = 300
 
 # Twitch channel for bot to join
+#UNIQUE#
 TWITCH_CHANNEL = "lunasparkai"
 
 # Voice reference file for TTS
+#UNIQUE#
 VOICE_REFERENCE = "neuro.wav"
 
 # MULTIMODAL SPECIFIC SECTION: Not needed when not using multimodal capabilities
@@ -34,12 +39,14 @@
 # LLM SPECIFIC SECTION: Below are constants that are specific to the LLM you are using
 
 # The model you are using, to calculate how many tokens the current message is
+# Ensure this is correct! Used for token count estimation
 MODEL = "meta-llama/Meta-Llama-3-8B"
 
 # Context size (maximum number of tokens in the prompt) Will target upto 90% usage of this limit
 CONTEXT_SIZE = 8192
 
 # This is your name
+#UNIQUE#
 HOST_NAME = "John"
 
 # This is the AI's name

main.py

@@ -46,7 +46,7 @@ def signal_handler(sig, frame):
     stt = STT(signals)
     # Create TTS
     tts = TTS(signals)
-    # Create LLMController
+    # Create LLMWrapper
     llm_wrapper = LLMWrapper(signals, tts, modules)
     # Create Prompter
     prompter = Prompter(signals, llm_wrapper)

memories/readme.md

@@ -5,7 +5,7 @@ will be automatically injected into the prompt. Memories will also persist acros
 the frontend or the database is deleted.
 
 The automatically generated memories are based off of 
-("Generative Agents: Interactive Simulacra of Human Behavior")[https://arxiv.org/abs/2304.03442]. Essentially, 
+[Generative Agents: Interactive Simulacra of Human Behavior](https://arxiv.org/abs/2304.03442). Essentially, 
 every handful of messages, the LLM will be prompted to review the recent messages and come up with the 3 most high level
 questions that encapsulate the conversation and also provide the answer. These question/answer pairs are then each
 stored as a (short-term) memory. These short-term memories will persists across restarts unless deleted.