yt-fts is a command line program that uses yt-dlp to scrape all of a YouTube
channels subtitles and load them into a sqlite database that is searchable from the command line. It allows you to
query a channel for specific key word or phrase and will generate time stamped YouTube urls to
the video containing the keyword.
It also supports semantic search via the OpenAI embeddings API using chromadb.
demo.mp4
pip
pip install yt-ftsDownload subtitles for a channel.
Takes a channel url as an argument. Specify the number of jobs to parallelize the download with the --jobs flag.
Use the --cookies-from-browser to use cookies from your browser in the requests, will help if you're getting errors
that request you to sign in. You can also run the update command several times to gradually get more videos into the database.
yt-fts download --jobs 5 "https://www.youtube.com/@3blue1brown"
yt-fts download --cookies-from-browser firefox "https://www.youtube.com/@3blue1brown"List saved channels.
The (ss) next to the channel name indicates that the channel has semantic search enabled.
yt-fts list┏━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ ID ┃ Name ┃ Count ┃ Channel ID ┃
┡━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ 1 │ ChessPage1 (ss) │ 19 │ UCO2QPmnJFjdvJ6ch-pe27dQ │
│ 2 │ 3Blue1Brown │ 127 │ UCYO_jab_esuFRV4b17AJtAw │
│ 3 │ george hotz archive │ 410 │ UCwgKmJM4ZJQRJ-U5NjvR2dg │
│ 4 │ The Tim Dillon Show │ 288 │ UC4woSp8ITBoYDmjkukhEhxg │
│ 5 │ Academy of Ideas (ss) │ 190 │ UCiRiQGCHGjDLT9FQXFW0I3A │
└────┴───────────────────────┴───────┴──────────────────────────┘
Full text search for a string in saved channels.
- The search string does not have to be a word for word and match
- Search strings are limited to 40 characters.
# search in all channels
yt-fts search "[search query]"
# search in channel
yt-fts search "[search query]" --channel "[channel name or id]"
# search in specific video
yt-fts search "[search query]" --video-id "[video id]"
# limit results
yt-fts search "[search query]" --limit "[number of results]" --channel "[channel name or id]"
# export results to csv
yt-fts search "[search query]" --export --channel "[channel name or id]" Advanced Search Syntax:
The search string supports sqlite Enhanced Query Syntax. which includes things like prefix queries which you can use to match parts of a word.
# AND search
yt-fts search "knife AND Malibu" --channel "The Tim Dillon Show"
# OR SEARCH
yt-fts search "knife OR Malibu" --channel "The Tim Dillon Show"
# wild cards
yt-fts search "rea* kni* Mali*" --channel "The Tim Dillon Show" You can enable semantic search for a channel by using the mbeddings command.
This requires an OpenAI API key set in the environment variable OPENAI_API_KEY, or
you can pass the key with the --openai-api-key flag.
Fetches OpenAI embeddings for specified channel
# make sure openAI key is set
# export OPENAI_API_KEY="[yourOpenAIKey]"
yt-fts embeddings --channel "3Blue1Brown"
# specify time interval in seconds to split text by default is 30
# the larger the interval the more accurate the llm response
# but semantic search will have more text for you to read.
yt-fts embeddings --interval 60 --channel "3Blue1Brown" After the embeddings are saved you will see a (ss) next to the channel name when you
list channels, and you will be able to use the vsearch command for that channel.
Starts interactive chat session with gpt-4o OpenAI model using
the semantic search results of your initial prompt as the context
to answer questions. If it can't answer your question, it has a
mechanism to update the context by running targeted query based
off the conversation. The channel must have semantic search enabled.
yt-fts llm --channel "3Blue1Brown" "How does back propagation work?"Summarizes a YouTube video transcript, providing time stamped URLS. Requires a valid YouTube video URL or video ID as argument. If the trancript is not in the database it will try to scrape it.
yt-fts summarize "https://www.youtube.com/watch?v=9-Jl0dxWQs8"
# or
yt-fts summarize "9-Jl0dxWQs8"output:
In this video, 3Blue1Brown explores how large language models (LLMs) like GPT-3
might store facts within their vast...
1 Introduction to Fact Storage in LLMs:
• The video starts by questioning how LLMs store specific facts and
introduces the idea that these facts might be stored in a particular part of the
network known as multi-layer perceptrons (MLPs).
• 0:00
2 Overview of Transformers and MLPs:
• Provides a refresher on transformers and explains that the video will focus
vsearch is for "Vector search". This requires that you enable semantic
search for a channel with embeddings. It has the same options as
search but output will be sorted by similarity to the search string and
the default return limit is 10.
# search by channel name
yt-fts vsearch "[search query]" --channel "[channel name or id]"
# search in specific video
yt-fts vsearch "[search query]" --video-id "[video id]"
# limit results
yt-fts vsearch "[search query]" --limit "[number of results]" --channel "[channel name or id]"
# export results to csv
yt-fts vsearch "[search query]" --export --channel "[channel name or id]"
Export search results:
For both the search and vsearch commands you can export the results to a csv file with
the --export flag. and it will save the results to a csv file in the current directory.
yt-fts search "life in the big city" --export
yt-fts vsearch "existing in large metropolaten center" --exportDelete a channel:
You can delete a channel with the delete command.
yt-fts delete --channel "3Blue1Brown"Update a channel: The update command currently only works for full text search and will not update the semantic search embeddings.
yt-fts update --channel "3Blue1Brown"Export all of a channel's transcript:
This command will create a directory in current working directory with the YouTube channel id of the specified channel.
# Export to vtt
yt-fts export --channel "[id/name]" --format "[vtt/txt]"