wsiAI

#!/usr/bin/zsh
#You can replace the shell above with bash, some effort is made to support both shells.
# Copyright (c) 2024 Quantius Benignus
# 
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
#--------------------------------------------------------------------------

# NAME: wsi 
# PREREQUSITES: 
#  - whisper.cpp installation (see https://github.com/ggerganov/whisper.cpp) or a whisperfile
#  - recent versions of 'sox', 'curl', 'xsel' or 'wl-copy' CLI tools from your system's repositories.
#  - optional llama.cpp or llamafile (and models) for AI assistant/translator functionality.
#  - optional xdotool (X11) or ydotool (Wayland) for automatic paste after successful transcription
#  - Piper neural text-to-speech engine for spoken AI assistant response 
#--------------------------------------------------------------------------
#X11 or Wayland (2nd line may catch edge cases):
wm="$XDG_SESSION_TYPE"
#wm="${XDG_SESSION_TYPE:-$(loginctl show-session $(loginctl | grep $(whoami) | awk '{print $1}') -p Type --value)}"

#---USER CONFIGURATION BLOCK----------------------------------------------------------------------
#Please, adjust the variables here to suit your environment:
# Store temp files in memory for speed and to reduce SSD/HDD "grinding":
TEMPD='/dev/shm'
# Hardcoded temp wav file to store the speech audio and get overwritten every time (in RAM):
ramf="$TEMPD/wfile"
#Set the number of processing threads for whisper.cpp inference (adjust for your case):
NTHR=8
#It seems that the optimum number of transcribe threads should equal half CPU processing cores:
#NTHR=$(( $(getconf _NPROCESSORS_ONLN) / 2 ))
# Use clipboard to paste from unless PRIMESEL set to true to use PRIMARY selection (middle mouse button to paste): 
PRIMESEL=false
# Try to paste automatically on completed transcription, reqires xdotool and such:
AUTOPASTE=true
#Provide hardcoded whisper.cpp hostname and port. To be used when invoked with -n option from CLI:
#(Can be overwritten when command line argument IP:PORT is supplied instead of '-n')
WHOST="127.0.0.1"
WPORT="58080"             

#The superdirectory where AI models (Whisper, LLMs, TTS etc.) are stored (adjust as needed):
AI="$HOME/REPLACE_W_PATH_TO/Models"
# Default whisper.cpp model file for local ASR inference (base.en provides excellent WER for English only)
#WMODEL="$TEMPD/ggml-base.en.bin"
WMODEL=${WHISPER_DMODEL:-"$AI/whisper/ggml-base.en.bin"}
# Uncomment to use available whisperfile (WITH BUILT-IN MODEL) instead of standalone whisper.cpp or whisper.cpp server
WHISPERFILE="whisper-tiny.en.llamafile" #available from https://huggingface.co/Mozilla/whisperfile/tree/main
#Use the following large language model with llama.cpp for the interactive assistant/translator etc.:
LLMODEL="$AI/gemma-2-9b-it-Q6_K_L.gguf" #avalable from https://huggingface.co/bartowski/gemma-2-9b-it-GGUF/tree/main   
#Use the following LIGHTWEIGHT large language model with llama.cpp for a simpler interactive assistant:
LIGHTLMODEL="$AI/gemma-2-2b-it-Q6_K_L.gguf" # available from https://huggingface.co/google/gemma-2-2b-it-GGUF
#Use the following heavy hitter for tasks that require more advanced reasoning (slower):
HEAVYMODEL="$AI/Qwen2.5-14B-Instruct-Q5_K_L.gguf" # available from https://huggingface.co/bartowski/Qwen2.5-14B-Instruct-GGUF
#Default llama.cpp executable:
llamf="llam" #This is an existing systemwide symbolic link to llama-cli from lllama.cpp
#Uncomment the following to use a LLM from a llammafile (also needs flag -a) instead of standalone llamma.cpp:
LLAMAFILE="$AI/gemma-2-2b-it.Q6_K.llamafile"
#LLAMAFILE="llamafile-0.8.16" #available from https://github.com/Mozilla-Ocho/llamafile/releases/tag/0.8.16
#Use the following text-to-speech Piper model for human-like audio response in English:
TTSMODEL="$AI/piper/en_US-lessac-low.onnx"  #available from https://github.com/rhasspy/piper/blob/master/VOICES.md
#The above TTS model has sample rate 16000 (if you change the model, adjust the sample rate below):
rtts="16000" 
#Use the following text-to-speech Piper model for human-like audio response in the language (e.g. chinese) of the LLM translator function:
TRANSMODEL="$AI/piper/zh_CN-huayan-medium.onnx"  #available from https://github.com/rhasspy/piper/blob/master/VOICES.md
#The above model has sample rate 22050 (if you change it, adjust the sample rate below):
rtrans="22050" 
#---END USER CONFIG BLOCK------------------------------------------------------------------------

#Notification code, prefers zenity, then notify-send, which should be available across the
#most distributions, in package libnotify or libnotify-bin:
desknote() {
    local title="$1"
    local message="$2"

    if command -v zenity &> /dev/null; then
        zenity --info --text="$message"
    elif command -v notify-send &> /dev/null; then
        notify-send "$title" "$message"
    elif command -v kdialog &> /dev/null; then
        kdialog --passivepopup "$message" 5
    else
        echo "Notification Message: $message" >&2
        echo "Please install zenity or notify-send to use notifications." >&2
        echo "You can install either using your package manager, e.g.:" >&2
        echo "  sudo apt-get install zenity or sudo apt-get install libnotify-bin" >&2
        echo "  sudo yum install zenity or sudo apt-get install libnotify" >&2
        echo "  sudo pacman -S zenity, etc." >&2
    fi
}

#---CHECK DEPENDENCIES. This block can be commented out once dependencies confirmed-----------------------------------
#The install-wsi script should have taken care of this, left here for the manual install.
command -v curl &>/dev/null || { echo "curl is required. Please, install curl" >&2 ; exit 1 ; }
#The next is only needed if curl requests json output from the whisper.cpp server;
#command -v jq &>/dev/null || { echo "jq is required. Please, install jq" >&2 ; exit 1 ; }
command -v sox &>/dev/null || { echo "sox is required. Please, install sox" >&2 ; exit 1 ; }
[ -n $WHISPERFILE ] || command -v transcribe &>/dev/null || { echo -e "Please, install whisper.cpp (see https://github.com/ggerganov/whisper.cpp)\
\nor download a whisperfile portable executable w model inside (see https://huggingface.co/Mozilla/whisperfile/tree/main)\
\nand create 'transcribe' in your PATH as a symbolic link to the chosen executable, e.g.\n \
'ln -s /full/path/to/whisper.cpp/main \$HOME/.local/bin/transcribe'" >&2 ; exit 1 ; }
command -v llam &>/dev/null || { echo -e "For interaction with an AI, install llama.cpp (see https://github.com/ggerganov/llama.cpp)\
\nor download a llamafile portable executable (see https://huggingface.co/Mozilla)\
\nand create in your PATH a symbolic link to the llama-cli executable, e.g.\n \
 'ln -s /full/path/to/llama.cpp/llama-cli \$HOME/.local/bin/llam'" >&2 ; exit 1 ; }
command -v piper &>/dev/null || { echo -e "Please, install piper text-to-speech (see https://github.com/rhasspy/piper)\
\nto use the AI assistant and translator features with human-like voice response." >&2 ; exit 1 ; }
# We will use zenity or notify-send (part of libnotify or libnotify-bin in some distros) for desktop notifications:
command -v zenity &>/dev/null || command -v notify-send &>/dev/null || { echo "zenity or notify-send needed for error reporting. Please, install zenity or libnotify-bin" >&2 ; exit 1 ; }
#Now let's check if we are in X11 or Wayland and use the right utility:
if [[ wm == "wayland" ]]; then
    command -v wl-copy &>/dev/null || { echo "wl-copy is needed for the clipboard. Please, install wl-copy" >&2 ; exit 1 ; } 
    if [ "$AUTOPASTE" = true ]; then 
       command -v ydotool &>/dev/null || { echo "To have the transcribed text pasted automatically at the cursor position, please install ydotool" >&2 ; exit 1 ; }
    fi
elif [[ wm == "x11" ]]; then
    command -v xsel &>/dev/null || { echo "We rely on xsel for the clipboard. Please, install xsel." >&2 ; exit 1 ; }
    if [ "$AUTOPASTE" = true ]; then 
       command -v xdotool &>/dev/null || { echo "To have the transcribed text pasted automatically at the cursor position, please install xdotool" >&2 ; exit 1 ; }
    fi
fi
#Check USER configuration vars:
for llv in $WMODEL $LLMODEL $LIGHTMODEL $HEAVYMODEL $TTSMODEL $TRANSMODEL
do
[[ -f $llv ]] || { desknote "File not found: " "$llv was not found. \nPlease, adjust in USER CONFIGURATION BLOCK." ; exi1 1 ; }
done 
command -v $LLAMAFILE &>/dev/null || { desknote "Not found: " "Executable $LLAMAFILE not found." ; exit 1 ; }
command -v $WHISPERFILE &>/dev/null || { desknote "Not found: " "Executable $WHISPERFILE not found." ; exit 1 ; }
echo -e "\nLooks like you have all dependencies installed.\nTo remove this message, comment out the 'CHECK DEPENDENCIES' block in the wsiAI script.\n"
desknote "Dependencies installed" "\nLooks like you have all dependencies installed.\n \
To remove this message, comment out the 'CHECK DEPENDENCIES' block in the wsiAI script.\n \
\nYou can run wsiAI --help from the Terminal to learn about its options..."
#---END CHECK DEPENDENCIES. The above block can be commented out after successful 1st run----------------------------

#Hear the complaints of the above tools and do not continue with the sequence:
set -e

# Process command line arguments
while [ $# -gt 0 ]; do
    case "$1" in
        -p|--primary)
            PRIMESEL=true
            shift
            ;;
        -c|--clipboard)
            # This is transitional warning due to breaking change (will be removed in the future). Clipboard is now the default:
            desknote "Breaking Change:" "The '-c' flag is deprecated, since clipboard is now default. \n \
            Remove unnecessary '-c' from your desktop shortcut definitions. \n \
            To use primary selection as the text source, use the '-p' flag in your desktop shortcuts."
            shift
            ;;
        -w|--whisperfile)
            whf="$WHISPERFILE"
            shift
            ;;
        -a|--llamafile)
            llamf="$LLAMAFILE --cli"
            shift
            ;;
        -n|--netapi)
            #This uses the hostname or IP and port specified in the config block
            #Can be overwritten, supplied as command line argument IP:PORT instead of this option
            IPnPORT="$WHOST:$WPORT" 
            if [[ "$(curl -s -f -o /dev/null -w '%{http_code}' $IPnPORT)" != "200" ]]; then
                echo "Can't connect to whisper.cpp server at provided address!" >&2
                desknote "No connection to Whisper.cpp server" "No whisper.cpp server at $IPnPORT."
                #Uncomment the next line if you want to have the server start locally on a call (assuming whserver symlink is set).
                #pidof whserver > /dev/null  2>&1 || { whserver --host 0.0.0.0 --port 58080 -t $NTHR -nt -m $WMODEL 2>/dev/null & }
                exit 1
            fi
            shift
            ;;
        -h|--help)
            echo -e "Usage: $0 [-p|--primary] [-n|--netapi] [-w|--whisperfile] [-l|--llamafile]\n"
            echo -e "  -p, --primary: Use PRIMARY selection instead of clipboard\n"
            echo -e "  -n, --netapi: Use whisper.cpp server with the host:port in the script's GONFIG block\n"
            echo -e "  -w, --whisperfile: Use whisperfile instead of standalone whisper.cpp. Note that the whisperfile will use the script-configured (not embedded) model.\n"
            echo -e "  -a, --llamafile: Use llamafile instead of standalone llama.cpp. Note that the llamafile will use the script-configured (not embedded) LLM model.\n"
            echo -e "Any other non-flag command-line argument is expected to be an <IP:port> pair. Error will occur if diferent."
            echo -e 'Speaking: "Assistant,..." followed by a question, will trigger the AI model response and that will be available for pasting'
            echo -e 'Speaking: "Computer,..." followed by a request, will make the LLM attempt to deliver a Linux command that matches the request'
            echo -e 'Speaking: "Coordinator,..." followed by a task, will make the LLM respond with the task formatted as a TODO item, as in Joplin inline TODOs'
            echo -e 'Saying: "Translator,..." followed by a statement, will make the AI model translate the statement and the translation willl be available for pasting'
            echo -e "The AI functions are experimental and work in progress." 
            echo -e "\nCAUTION: Transfering any AI-generated Linux commands to your command line is at your own risk. Make sure you know what you are doing."
            exit 0
            ;;
        *)
            #The network address and port should have already been sanitized
            IPnPORT=$1
            if [[ "$(curl -s -f -o /dev/null -w '%{http_code}' $IPnPORT)" != "200" ]]; then
                echo "Can't connect to a whisper.cpp server at provided address!" >&2   
                desknote "No connection to Whisper.cpp server" "No whisper.cpp server at $IPnPORT."
                #Uncomment the next line if you want to have the server start locally on a call (assuming whserver symlink is set).
                #pidof whserver > /dev/null  2>&1 || { whserver --host 0.0.0.0 --port 58080 -t $NTHR -nt -m $WMODEL 2>/dev/null & }
                exit 1
            fi
            shift
            ;;
    esac
done
#echo "Recording now: "
rec -q -t wav $ramf rate 16k silence 1 0.1 1% 1 2.0 3% channels 1 2>/dev/null

#Time stamps for timing the performance of the various stages (Un/Comment as needed): 
#stmp=$(date +%s%N)

if [ -n "$IPnPORT" ] ; then 
    #echo "Sending to server now: "
    str=$(curl -S -s $IPnPORT/inference \
        -H "Content-Type: multipart/form-data" \
        -F file="@$ramf" \
        -F temperature="0.0" \
        -F temperature_inc="0.2" \
        -F response_format="text")
elif [[ "$whf" == *.llamafile ]]; then
    #echo "Using $whf:"
    str="$($whf -t $NTHR -nt --gpu auto -f $ramf 2>/dev/null)"     
else
    #A default fallback if the whisper server is offline.
    #echo "Transcribing now: "
    str="$(transcribe -t $NTHR -nt -fa -m $WMODEL -f $ramf 2>/dev/null)" 
fi

#echo "Got text back after: "$((($(date +%s%N) - stmp)/1000000))" ms" 
#stmp=$(date +%s%N)

#Please, note, that if you have disabled extended globbing for some reason, you likely know how to enable it only for this script. 
# Whisper detected non-speech events such as (wind blowing): 
str="${str/\(*\)}"   
str="${str/\[*\]}"
str="${str#$'\n'}"    
str="${str#$'\n'}"
#Prefer the power of zsh, but loose full POSIX compliance.
if [ -n "$ZSH_NAME" ]; then
   str="${str# }"
   str="${str# }"
   str="${(C)str:0:1}${str#?}"
  #Are we engaging with a LLM via a 'wake' word?
   if [[ "$str" == Assistant*  ]]; then
     str="${str#Assistant[,.] }"
     str="${str%'\n'}"
     strai=$($llamf -t $NTHR -c 2048 --temp 0 -ngl 99 --no-display-prompt -n 150 -m $LIGHTLMODEL --prompt "Provide a concise response (but do not explain why is it concise) to the following: $str" 2>/dev/null )
     strai="${strai/\[end of text\]}"
     echo "$strai" | sed 's/\*//g' | piper --model $TTSMODEL --output-raw 2>/dev/null | aplay -r $rtts -f S16_LE -t raw - 2>/dev/null
     str="${(C)str}\n$strai" #We can still paste the result
   elif [[ "$str" == [Tt]ranslator*  ]]; then
     echo $str
     str="${str#[Tt]ranslator[,. ] }"
     echo "$str" | piper --model $TTSMODEL --output-raw 2>/dev/null | aplay -r $rtts -f S16_LE -t raw - 2>/dev/null
     $llamf -t $NTHR -c 1024 --temp 0 -ngl 99 -n 80 -m $LLMODEL --no-display-prompt --prompt "Respond only with the chinese translation, written in chinese only, and nothing more. Translate the following to chinese: $str" 2>/dev/null | sed 's/\*//g; s/\[end of text\]//g' | tee /dev/tty /dev/shm/lmf | piper --model $TRANSMODEL --output-raw 2>/dev/null | aplay -r $rtrans -f S16_LE -t raw - 2>/dev/null
     str="${$(< /dev/shm/lmf)//$'\n'/}"
   elif [[ "$str" == [Cc]omputer*  ]]; then
     if [[ $str == *Grammarly* || $str == *proofread* ]]; then
        if [[ wm == "x11" ]]; then
            selpaste="$(xsel --primary --output)"
        elif [[ wm == "wayland" ]]; then
            selpaste="$(wl-paste --primary --type text)"
        else
            #X11 fallback
            selpaste="$(xsel --primary --output)"
        fi
        str='<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n<|im_start|>user\nBehave like Grammarly on this, by simply providing only the corrected text and nothing else, no explanations, no markdown, etc.: '
        strai=$($llamf -t $NTHR -c 4096 --temp 0 -ngl 99 -mli --no-display-prompt -m $HEAVYMODEL --prompt "$str$selpaste<|im_end|>\n<|im_start|>assistant\n" 2>/dev/null )
        strai="${strai//\*/}"
        str="${strai/\[end of text\]}"
     else
        strai=$($llamf -t $NTHR -c 256 --temp 0 -ngl 99 --no-display-prompt -n 100 -m $LLMODEL --prompt "You are a Linux command helper. To any request, you respond only with a precise linux command. Do not output any other text. Here is the request: $str" 2>/dev/null )
        # | sed 's/[`\*]//g' )
        strai="${strai//\*/}"
        str="${strai/\[end of text\]}"
     fi
   elif [[ "$str" == [Cc]oordinator*  ]]; then
     strai=$($llamf -t $NTHR -c 512 --temp 0 -ngl 99 --no-display-prompt --simple-io -n 100 -m $LLMODEL --prompt "You are a helpful scheduler. When I ask you to schedule a task, you will format it like this: - [ ] @TODO Task body. Let's give it a try: $str" 2>/dev/null )
     str="${strai/\[end of text\]}"
   fi
elif [ -n "$BASH" ]; then
   #Running in bash because you changed the shebang on line 1
   str="${str#*(+([![:space:]]))}"
   str="${str^}"
  #Are we engaging with a LLM via a 'wake' word?
   if [[ "$str" == ?[Aa]ssistant* ]]; then
     str="${str#?[Aa]ssistant[,.] }"
     str="${str^}"
     $llamf -t $NTHR -c 2048 --temp 0 2>/dev/null -ngl 999 -m $LIGHTLMODEL --prompt "Provide a very concise response to the following: $str"  --no-display-prompt -n 150 | tee /dev/tty /dev/shm/lmf | sed 's/\*//g' | piper --model $TTSMODEL --output-raw 2>/dev/null | aplay -r $rtts -f S16_LE -t raw - 2>/dev/null
     str="$str\n"$(< /dev/shm/lmf) #We can still paste the result
   elif [[ "$str" == ?[Tt]ranslator* ]]; then
     str="${str#?[Tt]ranslator[,.] }"
     echo "$str" | piper --model $TTSMODEL --output-raw 2>/dev/null | aplay -r $rtts -f S16_LE -t raw - 2>/dev/null
     $llamf -t $NTHR -c 2048 --temp 0 2>/dev/null -ngl 999 -m $LIGHTLMODEL --no-display-prompt --prompt "Respond in chinese and be concise. Translate the following to chinese: $str"  -n 80 | tee /dev/tty /dev/shm/lmf | sed 's/\*//g' | piper --model $TRANSMODEL --output-raw 2>/dev/null | aplay -r $rtrans -f S16_LE -t raw - 2>/dev/null
     str="$(< /dev/shm/lmf)"
     str=${str#*(+([![:space:]]))}
   elif [[ "$str" == [Cc]omputer*  ]]; then
     if [[ $str == *Grammarly* || $str == *proofread* ]]; then
        if [[ wm == "x11" ]]; then
            selpaste="$(xsel --primary --output)"
        elif [[ wm == "wayland" ]]; then
            selpaste="$(wl-paste --primary --type text)"
        else
            #X11 fallback
            selpaste="$(xsel --primary --output)"
        fi
        str='<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n<|im_start|>user\nBehave like Grammarly on this, by simply providing only the corrected text and nothing else, no explanations, no markdown, etc.: '
        strai=$($llamf -t $NTHR -c 4096 --temp 0 -ngl 99 -mli --no-display-prompt -m $HEAVYMODEL --prompt "$str$selpaste<|im_end|>\n<|im_start|>assistant\n" 2>/dev/null )
        strai="${strai//\*/}"
        str="${strai/\[end of text\]}"
     else
        strai=$($llamf -t $NTHR -c 256 --temp 0 -ngl 99 --no-display-prompt -n 100 -m $LLMODEL --prompt "You are a Linux command helper. To any request, you respond only with a precise linux command. Do not output any other text. Here is the request: $str" 2>/dev/null )
        # | sed 's/[`\*]//g' )
        strai="${strai//\*/}"
        str="${strai/\[end of text\]}"
     fi
   elif [[ "$str" == [Cc]oordinator*  ]]; then
     strai=$($llamf -t $NTHR -c 512 --temp 0 -ngl 99 --no-display-prompt -n 100 -m $LLMODEL --prompt "You are a helpful scheduler. When I ask you to schedule a task, you will format it like this: - [ ] @TODO Task body. Let's give it a try: $str" 2>/dev/null )
     str="${strai/\[end of text\]}"
   fi
else
    #Not testing for AI input if shell is unknown:
    echo "Unknown shell, assuming bash compliance"
    str="${str##+([[:space:]])}"
    str="${str^}"
fi

#We have a result (either recognized text or AI response), now we make a few decisions:
#If this is somehow run in a text console: 
if [[ -z "${DISPLAY}" ]] || [[ -z "${DESKTOP_SESSION}" ]] || [[ -z "${XDG_CURRENT_DESKTOP}" ]]; then
#"Not running in a known graphics environment. Using standard output:
    echo $str ; exit 0
else
#xsel or wl-copy:
#echo "Pasting now: "
 case "$wm" in
    "x11")
        if [ "$PRIMESEL" = true ]; then
          echo $str | xsel -ip
          # Simply clicking the middle mouse button. The user has to take care of: 1. Window choice. 2. Position within window.
          # Thus, autopaste from the primary selection is more unpredictable and requires extra care of window focus and mouse pointer repositioning. 
          if [ "$AUTOPASTE" = true ]; then xdotool click 2 ; fi
        else
          echo $str | xsel -ib
          # The automatic paste option makes more sense when using the CLIPBOARD, not primary sellection.
          if [ "$AUTOPASTE" = true ]; then xdotool key ctrl+v ; fi
        fi
        ;;
    "wayland")
        if [ "$PRIMESEL" = true ]; then
          echo $str | wl-copy -p
          # Simply clicking the middle mouse button. The user has to take care of: 1. Window choice. 2. Position within window.
          # Thus, autopaste from the primary selection is more unpredictable and requires extra care of window focus and mouse pointer repositioning. 
          if [ "$AUTOPASTE" = true ]; then ydotool click 0xC2; fi
        else
          echo $str | wl-copy
          # The key sym sequence may differ from the one below on a variety of systems and keyboard layouts:
          if [ "$AUTOPASTE" = true ]; then ydotool key 37:1 55:1 55:0 37:0 ; fi
        fi 
        ;;
    *)
        echo $str
        ;;
 esac
fi

#echo "Text processed after another: "$((($(date +%s%N) - stmp)/1000000)) "ms"