Video Search API BETA

Find the scenes your models need. Search 1B+ indexed web
videos by visual content, not just metadata, and get
timestamped results that feed directly into your training-data
pipeline. Built for VLM, VLA, and world model data curation at
web scale.

Request beta access
Access is gated by a KYC (Know Your Customer) review to ensure compliant, documented use.
  • 1B+ web videos indexed by visual content
  • Visual query & text-to-video search
  • Timestamped, moment-level results
  • Public dataset indexing on request

Powering AI training pipelines on a platform trusted by 50,000+ customers

Foundation model labs
Robotics & VLA teams
AV programs
Video generation companies
Dataset curators
1B+
Web videos indexed
10B+
Video extractions/day platform capacity
100%
Pipeline-native: feeds the Media Data API

Web-scale visual discovery for AI training

Keyword search finds titles and tags. Video Search API finds the visual content itself: the scenes, moments, and task families your models actually need.

Visual-First Search

Submit an image, frame, or text description of a target scene. Get ranked video results that match by visual content, surfacing moments never described in titles or tags.

Timestamp-Level Results

Results point to specific moments, not just video-level metadata. Jump straight to the frames that matter for your training set.

Training-Pipeline Native

Output feeds directly into the Media Data API for retrieval. Discovery and extraction on the same platform.

Search by task, not by title

One API call turns a visual query or natural-language description into ranked, timestamped video results, ready for your extraction pipeline.

                              curl -sX POST https://api.brightdata.com/video-search/v1/text 
  -H "Authorization: Bearer " 
  -H "Content-Type: application/json" 
  -d '{"query":"first-person view washing
 dishes","date_from":"2026-03-01","date_to":"2026-03-17","top_k":20,"min_score":0.20}'
From query to curated dataset
  • Query : submit a visual example or a text description of a target scene
  • Discover : get ranked results with source URLs and timestamps
  • Retrieve : pipe output into the Media Data API for extraction
  • Augment : point us at public video datasets of interest and we'll index them
Request beta access

Built for systematic training-data curation

1B+ Video Web Index
Pre-indexed web videos searchable by visual content. A live index, not a frozen academic dataset.
Visual Query Support
Submit an image or frame to find visually similar scenes across the entire index.
Text-to-Video Search
Natural-language descriptions return ranked, timestamped results.
Timestamp-Level Results
Results point to specific moments and scenes, not just video-level metadata.
Public Dataset Indexing
Interested in a specific public video dataset? Point us at it and we'll index it alongside the web index.
Media Data API Integration
Output feeds directly into the Media Data API for retrieval.
VLA Task-Family Queries
Designed for task-family discovery: "pick and place," "door opening," and beyond.
Continuously Refreshed
Index is updated as part of Bright Data's ongoing web crawl cadence.

Video Search API

The only web-scale visual discovery index for AI training workflows, enabling systematic task-family discovery and scene diversity at a scale no manual curation team can match.

Request beta access

Power your most demanding visual workflows

VLM & VLA training data curation
Surface specific scenes and moments for vision-language and vision-language-action model training datasets.
World model training
Build the real-world scene diversity that world models require. Robotics and AV programs use targeted visual queries to cover environments, dynamics, and edge cases that static datasets miss.
Video generation reference discovery
Find reference clips matching target aesthetics, motion styles, or scene compositions.
Content moderation at scale
Visual discovery of policy- violating content patterns across web video without manual review.
Request beta access
Video Search API vs keyword video search
Recommended

Video Search API

Web-scale visual discovery
Matching
Visual content matching that finds scenes never described
Index
1B+ live web videos, continuously updated
Results
Timestamped, moment-level, ranked by visual relevance
Pipeline
API output designed for ML pipelines; feeds the Media Data API
Custom coverage
Public dataset indexing on request.

Keyword search / static datasets

YouTube search, WebVid, Panda-70M
Matching
Keyword-only matching misses the long tail of relevant scenes
Index
Academic datasets are frozen, taxonomy-constrained, and outdated
Results
Video-level metadata only, no moment-level targeting
Pipeline
No API for bulk discovery; no training-pipeline integration
Custom coverage
No way to add coverage of the public datasets you care about

Find the scenes your models need

Web-scale visual discovery for AI training, now in Beta.

Request beta access

FAQs

Video Search API indexes web videos to enable discovery of specific frames, moments, and scenes. Given a visual query or text description, it returns ranked video results with timestamps pointing to relevant segments. Built for ML teams curating training data for VLMs, VLAs, and world models.

Keyword search finds metadata. Video Search API finds visual content: scenes where the relevant moment isn't described in the title or tags. That's where task-family diversity comes from, and it's why keyword search can't scale to frontier model training requirements.

No. Twelve Labs searches and understands content you already hold. Video Search API discovers new content from the 1B+ web index. They are complementary products targeting different workflow steps. Use Video Search API to discover, and an understanding layer to process.

Yes, as long as the videos are public. Point us at public video datasets you're interested in and we'll index them, so you can query them alongside the web index. Private datasets are not supported.

Output currently feeds into the Media Data API for retrieval after discovery. Additional integrations may follow after Beta.

Visual content only. Audio search is not currently supported.

The index is updated as part of Bright Data's ongoing web crawl cadence. Frequency is subject to Beta SLA terms.

Beta means we are still improving the search algorithm, and result quality will keep getting better. The underlying index and infrastructure are production-grade, built on the same platform that supports 10B+ video extractions per day. Ask about current beta customer results.

KYC (Know Your Customer) is our vetting process: describe your intended use case, data types, and end purpose. Bright Data approves access based on compliance with the acceptable use policy and applicable law. Output includes source URL and timestamp; rights determination is the customer's responsibility.