flashdata-dev's picture
Publish youtube_transcript tool and request guide
87b4224 verified
|
Raw History Blame
3.96 kB
---
title: "YouTube transcripts for cited research"
emoji: 🔎
colorFrom: blue
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
license: mit
tags:
- smolagents
- tool
- flashdata
- youtube-transcript
---
# YouTube transcripts for cited research — FlashData
Give a research agent timestamped text from an existing YouTube caption track. Keep segment start and end times so a reader can return to the video when checking a generated summary.
**Try this task:** Read existing English automatic captions for a timestamped video brief.
[FlashData youtube transcript documentation](https://flashdata.dev/docs/sources/youtube_transcript?utm_source=hugging-face&utm_medium=integration&utm_campaign=intent_gtm_2026q4&utm_content=youtube-transcript-research-flashdata) · [Create a FlashData account](https://flashdata.dev/signup)
## Run it
Open the Space to edit a request and inspect a documentation response sample. The interactive page runs locally in your browser; it does not send a FlashData query or collect an API key. The Python tool runs in your own process.
Use Python 3.10 or newer. Install the pinned dependencies:
```sh
python -m pip install "smolagents==1.26.0" "langchain-flashdata==0.2.1"
```
Create an API key in the FlashData console, fund the required credits, and supply the key in your local environment. Review `tool.py` before enabling remote code execution.
```python
import getpass
import os
from smolagents import load_tool
os.environ["FLASHDATA_API_KEY"] = getpass.getpass("FlashData API key: ")
tool = load_tool("flashdata-dev/youtube-transcript-research-flashdata", trust_remote_code=True)
result = tool(**{'auto': True, 'language': 'en', 'video_id': 'FUfGcZ092b0'})
print(result)
```
This is a native smolagents Tool with typed task inputs. Add the loaded tool to your own agent; model choice, prompts and model fees remain under your control. Direct calls above do not need a model.
## Inputs
| Parameter | Type | Required / default | Details |
| --- | --- | --- | --- |
| `video_id` | string | Required | 11-character YouTube video ID or a YouTube video URL. |
| `language` | string | 'en' | Requested caption language code. pattern=^[A-Za-z0-9-]{1,35}$ |
| `auto` | boolean | True | True requests automatic captions; false requests manually authored captions. |
Optional `None` values are omitted; the documented adapter defaults apply.
Full validation rules are in `input-schema.json`.
## Output and next step
The useful data is at `results.0.segments`. Relevant fields: `start`, `end`, `text`.
Retain video ID, language and timestamps when passing segments to your own model. Model calls and summarization are separate steps.
The Space includes a recorded documentation example and its original request. It is not a fresh result for the task prompt above. Optional fields may be absent.
## Costs, limits and failures
Defaults to existing automatic English captions. No ASR, translation or automatic fallback to another language or caption source. Missing tracks fail.
The tool is MIT licensed. It is free to install. **FlashData API calls are paid**; installation does not include free data queries. Check current [FlashData pricing](https://flashdata.dev/pricing) and your console balance. Platform, model and customer storage charges are separate. Each search/page is a separate request; Extended Search has its own rate. Downloads are charged by delivered data volume.
An uncertain timeout can still represent a charged request. Check Jobs in the FlashData console before trying again; this integration does not automatically retry paid submissions. A 401 means the key was rejected, 402 means insufficient credits, and 429 means a rate or quota limit.
[Get started with this FlashData capability](https://flashdata.dev/docs/sources/youtube_transcript?utm_source=hugging-face&utm_medium=integration&utm_campaign=intent_gtm_2026q4&utm_content=youtube-transcript-research-flashdata)