langchain-transcriptfetch

A LangChain document loader for TranscriptFetch: transcripts of YouTube, TikTok and Instagram videos, from a single link to whole channels, playlists and searches, split server-side into embedding-sized chunks with a link to the moment each one starts.

pip install -U langchain-transcriptfetch
export TRANSCRIPTFETCH_API_KEY="tf_live_..."   # free key: https://transcriptfetch.com

Load a few videos

from langchain_transcriptfetch import TranscriptFetchLoader

docs = TranscriptFetchLoader([
    "https://www.youtube.com/watch?v=aircAruvnKk",
    "https://www.tiktok.com/@user/video/7428000000000000000",
]).load()

One Document per video. metadata holds source (the video URL), title, channel, platform, duration, language and transcript_source (captions or audio).

A whole channel, chunked for RAG

docs = TranscriptFetchLoader(channel="@3blue1brown", max_videos=20, chunks=True).load()

docs[0].metadata
# {'source': 'https://www.youtube.com/watch?v=aircAruvnKk&t=4s',
#  'video_url': 'https://www.youtube.com/watch?v=aircAruvnKk',
#  'chunk_index': 0, 'start': 4.22, 'end': 114.23, 'tokens': 400,
#  'title': 'But what is a neural network? | Deep learning chapter 1', ...}

With chunks=True the transcript is split by the API between caption lines (never mid-word), with a small overlap, so you can skip the text splitter. source links to the exact moment, which makes citations in answers clickable. Tune the size with chunk_tokens (100 to 2000, default 500) and chunk_overlap (0 to 200, default 50).

from langchain_core.vectorstores import InMemoryVectorStore
from langchain_openai import OpenAIEmbeddings

store = InMemoryVectorStore.from_documents(docs, OpenAIEmbeddings())
hits = store.similarity_search("what does a neuron hold?", k=3)
for h in hits:
    print(h.metadata["source"], h.page_content[:80])

Sources

Pass exactly one:

Listings are capped at max_videos (default 25).

Options

• mode="auto" (default) reads captions and transcribes the audio when a video has none, which most TikToks and Reels need. mode="captions" never transcribes, so captionless videos are skipped and cost nothing.
• wait_for_audio=True (default) waits for audio transcriptions to finish (audio_timeout, default 900 seconds each).
• continue_on_failure=True (default) logs and skips a video that fails instead of raising.
• api_key and base_url override the environment.

lazy_load() streams Documents as they arrive. Transcripts are fetched 50 at a time.

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. See pricing.

License

MIT