qa.sh

#!/bin/sh

# First argument is video URL or ID
# Second argument is the question

# set -uo pipefail

# Get video ID from URL
VIDEO_ID=$(echo $1 | cut -d'=' -f 2)

# Get question from args
QUESTION=$2

AUDIO="whisper_$VIDEO_ID.wav"

TITLE_FILE="$VIDEO_ID.title"
DESCRIPTION_FILE="$VIDEO_ID.desc"
TAGS_FILE="$VIDEO_ID.tags"
UPLOADER_FILE="$VIDEO_ID.uploader"
JSON_FILE="$VIDEO_ID.info.json"
TEXT_FILE="$AUDIO.txt"

# Download audio
if [ ! -e $AUDIO ]; then
    yt-dlp --write-info-json -f bestaudio --extract-audio $1 -o "%(id)s.%(ext)s"
    # Convert video to 16
    ffmpeg -i $VIDEO_ID.opus -f wav -ar 16000 $AUDIO
fi

# Create prompt
UPLOADER=$(cat $JSON_FILE | jq -r '.uploader')
TITLE=$(cat $JSON_FILE | jq -r '.title')
DESCRIPTION=$(cat $JSON_FILE | jq -r '.description')
TAGS=$(cat $JSON_FILE | jq -r '.tags[]')

PROMPT=$(cat << EOF
Please transcribe the following video using the following meta data for your vocabulary.

Title: $TITLE
Uploader: $UPLOADER
Tags: $TAGS
Description:
$DESCRIPTION
EOF
)

# Remove URLs from prompt
PROMPT=$(echo $PROMPT | sed -E 's/http[^ ]+//g')

# # Transcribe audio
if [ ! -e $TEXT_FILE ]; then
    ../whisper.cpp/main \
    -m ../whisper.cpp/models/ggml-large.bin \
    -l auto \
    -tr \
    --prompt "$PROMPT" \
    -otxt \
    $AUDIO
fi

# define multline string variable using EOF

# To add in system or user prompt:
#
# - you must only use information provided in the transcript
# - you may also rely on common knowledge
# - if you rely on specialized knowledge that does not appear in the transcript, you must explicitly say so

GPT_PROMPT=$(cat << EOF
# Video meta data

Title: $TITLE
Uploader: $UPLOADER
Tags: $TAGS
Description:
$DESCRIPTION

# Video transcript

$(cat $TEXT_FILE)

# Instructions

$QUESTION

Please follow the instructions (if unclear, assume answering a question) based on the youtube video transcript and meta data.
EOF
)

# Send transcript to LLM along with question
pipx run llm -m gpt4 "$GPT_PROMPT"