Oracle builds natural-language search for Red Bull Racing’s video archive

Finding a podium celebration or a sponsor’s logo in years of racing footage need not start with knowing the right filename. Oracle said on October 30 that it has developed a natural-language search solution for Oracle Red Bull Racing’s video archive, designed to take media staff from a plain-English request to relevant, playable clips.

The immediate problem is production work, not race strategy. According to Oracle, the team’s media staff currently browse folders, filenames and manually applied tags. Finding a specific moment requires knowing where to look, then reviewing the asset to locate the clip. That limits how many searches staff can complete during time-sensitive media and sponsor requests.

The disclosure describes a system developed to evaluate whether searching by meaning can improve that process. It does not establish a completed rollout or provide measured search-performance results. Oracle identifies latency, accuracy, capacity and operational value as measures to compare against the manual workflow.

Describe the footage once, search it live

The design separates preparation from searching. A background pipeline processes selected footage, while a live search service handles requests. Oracle says that separation keeps search available as new video is processed. Raw footage stays in Oracle Cloud Infrastructure’s Object Storage; the searchable information goes into Autonomous AI Lakehouse, Oracle’s database platform.

NVIDIA GPUs accelerate the preparation work: generating frame descriptions, detecting sponsor logos and tagging key moments. Face recognition can match approved people against an enrolled gallery. Oracle’s Private AI Services Container then turns generated captions into embeddings—numerical representations used to match content by meaning. Captions, embeddings, people information and sponsor evidence are stored together.

At search time, the database creates an embedding for the user’s question. Oracle AI Vector Search finds matching clips, reranks them and removes duplicates. The intended result is a jump to the relevant timecode, rather than another file that someone must watch from the beginning.

The database decides which clips can be returned

Oracle keeps the vectors alongside metadata and user permissions rather than moving them into a separate vector database. Retrieval uses SQL, the language for querying database records, to combine meaning-based matches with structured filters and access rules. Oracle says the resulting answers return only evidence the user is authorized to see.

Beyond locating moments, Oracle describes three uses for the same enriched footage:

• Find recognized people through face matches, rather than merely returning people who look similar.
• Measure sponsor exposure through on-screen minutes and each brand’s share of visibility.
• Produce recaps and answers with citations linking back to source clips.

The full archive remains a capacity question

The size of the searchable collection is not fully clear. Oracle’s write-up describes 22 TB of archived video assets, but separately estimates the ten-season archive at 80–90 TB. It does not explain the relationship between those figures, so neither establishes how much footage has been processed for search.

Scaling also has a specific memory constraint. The search indexes described by Oracle are held in memory, and grow with the number of vectors and the chosen configuration. The company identifies several ways to manage expansion: add database resources, divide workloads, adjust sampling or numerical precision, and accelerate index building with GPUs where that proves useful.