LlamaIndex Readers Integration: TranscriptFetch

Transcripts of YouTube, TikTok and Instagram videos as LlamaIndex Documents, from a single link to whole channels, playlists and searches, via TranscriptFetch. With chunks=True the API splits each transcript into embedding-sized pieces with a link to the moment each one starts, so answers can cite the exact second.

pip install llama-index-readers-transcriptfetch
export TRANSCRIPTFETCH_API_KEY="tf_live_..."   # free key: https://transcriptfetch.com

Usage

from llama_index.readers.transcriptfetch import TranscriptFetchReader

reader = TranscriptFetchReader()
docs = reader.load_data(["https://www.youtube.com/watch?v=aircAruvnKk"])

A whole channel, chunked, into an index:

from llama_index.core import VectorStoreIndex

reader = TranscriptFetchReader(chunks=True)
docs = reader.load_data(channel="@3blue1brown", max_videos=20)

index = VectorStoreIndex.from_documents(docs)
response = index.as_query_engine().query("What does a neuron hold?")
for node in response.source_nodes:
    print(node.metadata["source"])  # https://www.youtube.com/watch?v=...&t=4s

Each chunk Document carries source (a link to that moment), video_url, chunk_index, start, end, tokens, title, channel and platform. Links, IDs and timings are kept out of the embedded text.

Sources

load_data() takes exactly one: videos (a URL or ID, or a list: YouTube, TikTok, Instagram, or a direct audio or video file URL), channel (a YouTube @handle, URL or ID, or a TikTok or Instagram profile URL), playlist, or search (YouTube unless search_platform="tiktok" or "instagram"). Listings are capped at max_videos (default 25). lazy_load_data() streams instead.

Options

Set on the reader: mode ("auto" transcribes the audio of videos without captions; "captions" skips them and costs nothing), chunks, chunk_tokens (100 to 2000, default 500), chunk_overlap (0 to 200, default 50), wait_for_audio, audio_timeout, continue_on_failure, api_key, base_url.

Billing

A transcript or a listing page costs 1 credit, audio transcription 1 credit per started 5 minutes of audio, and failed requests are free. Every account gets 50 free credits a month, no card needed.