How to use Shade's AI tagging and semantic search

Shade's AI works best when you stop treating it like a search box and start treating it like an assistant you have to brief. The auto-tagging, the semantic ("neural") search, and the face clustering are all genuinely good, but the difference between a library that answers "find every reaction shot of the CEO at dusk" in one second and a library that returns junk comes down to a few setup choices you make in the first hour. This is a practical guide to getting those choices right: what Shade indexes, how to write custom metadata prompts that actually land, how face labeling works, and where the AI quietly falls short so you can plan around it. Prices and specs below were checked Jun 2026.

What Shade actually indexes, and where it runs #

When a file lands in Shade, three things happen automatically at ingest. The clip gets transcribed with timecodes, the visuals get described scene by scene, and faces get detected and grouped into clusters. All three feeds pour into one search index, which is why a single plain-English query can pull from spoken words, on-screen content, and people at the same time. Think of it less like tagging a photo and more like having a researcher watch every clip and take notes you can later query.

Two architecture facts shape everything else. First, this runs in Shade's cloud, not on your machine. ShadeFS, the FUSE-based mount that launched in January 2026, lets you open a 4K or 8K file in Premiere or DaVinci without a full download, but the AI indexing itself happens server-side after upload (checked Jun 2026). Second, Shade splits the AI workload: it uses OpenAI for custom metadata generation, and it distilled larger open-source vision models in-house so its semantic search runs fast on CPU rather than expensive GPUs across hundreds of thousands of assets. Transcription is handled by AssemblyAI under the hood. That hybrid is the reason indexing a big library is cheap enough to be "unlimited" on the paid plans, which I will come back to.

Auto-tagging and custom metadata fields #

Out of the box Shade generates scene descriptions, shot detection, transcripts, and IPTC/EXIF extraction with no setup. The real power, though, is custom metadata fields you define yourself. You create these under Settings, then your drive, then Metadata. You give the field a name, pick a type (text, single-select, or tags), choose which asset types it applies to, and then toggle "Autofill with AI" and write a prompt telling the model what to look for. For select fields you can also let the AI generate new options it has not seen before.

The prompt is where good results are won or lost, and Shade's own examples are instructive because they are blunt and specific:

The pattern that works: tell the model the exact decision you want it to make, give it the vocabulary when you have one, and keep each field doing one job. Teams have trained Shade to recognize remarkably narrow things this way, from specific paver types and crane models on a construction shoot to distinguishing sports teams by uniform. The mistake I see is writing a single vague "describe this" field and hoping it covers everything. Split it. A "shot type" select, a "time of day" select, and a "products visible" tag field will each be more accurate than one prompt asked to do all three.

Semantic search: searching by meaning, not filename #

This is the feature Shade leads with, and it earns the billing. At NAB 2026, co-founder Brandon Fan searched "a person skiing while holding a laptop" and the system jumped straight to the timestamped moment inside a longer clip where that scene appears, rather than just surfacing the file (checked Jun 2026). That sub-clip precision is the part that changes your day. You are not opening five files and scrubbing; you land on the frame.

To get good results, lean into descriptive, natural phrasing. "People running along the lake at sunset" beats "run lake." Because the index fuses transcript, scene description, and face data, you can also combine modalities in one query: a spoken phrase plus a visual plus a person. The honest caveat, from independent reviews, is that search quality on niche or poorly transcribed footage is still unproven at scale. Shade launched publicly in early 2025 and reported roughly 94 paying customers as of the 2026 reviews, so the long tail of weird, jargon-heavy, or badly miked footage has not been stress-tested across thousands of libraries yet. For mainstream content it is strong. For a heavily technical or noisy archive, run a real test on your own footage before you commit your whole catalog.

Face clustering and labeling people #

Shade detects faces at ingest and groups them into clusters automatically, then lets you put a name on a cluster once. After that you can search "all clips with Sarah" or "every reaction shot of the CEO" and get matching takes back regardless of folder, project, or filename. The company built its facial recognition in-house specifically because off-the-shelf clustering choked at the scale a real media library demands.

The workflow that gets you clean results: label the recurring people first, before you go searching by name. The cluster has to exist and be named for name search to work, so a few minutes naming your principal cast, hosts, or executives up front pays off across the whole library. One real-world note for anyone in regulated or sensitive work: people-categorization is powerful and also a liability surface, because once footage is searchable by individual, "who appears in this" becomes a query anyone with access can run. I wrote about whether that is useful or a risk in facial recognition and people-categorization in creative tools, and it is worth a read before you turn it on for embargoed or NDA footage.

Privacy, cost, and the cloud tradeoff #

Because all of this runs in Shade's cloud, two questions matter: who can see your data, and what does it cost. On the first, Shade publishes an AI Privacy Guarantee stating your data is 100% your property and is never used to train its models, and it holds SOC 2 Type II, ISO 27001, HIPAA, GDPR, and TPN certifications, with end-to-end encryption in transit and at rest (checked Jun 2026). As with any vendor-stated guarantee, ask for the audit reports directly if your clients require it. The broader "should I send footage to a cloud AI at all" question I cover in the privacy cost of cloud AI search.

The "unlimited AI indexing" line is real and is the best part of the pricing: the distilled CPU search engine is what makes it affordable to index everything rather than rationing it. The catch is the per-seat storage math. Active storage is capped per seat (500 GB on Growth, 1 TB on Enterprise), so a team with a large hot working set adds up faster than the sticker price implies, which is exactly how the $20 seat becomes a five-figure annual contract.

This is the one honest place JuiceMount enters the conversation, and only because the comparison is native to the topic. JuiceMount is an open-source, $0-per-seat mount layer over your own NAS, and its search index runs locally rather than in a vendor cloud. That suits teams who want to keep footage on-prem and skip per-seat AI fees. Where it does not fit: JuiceMount does not ship Shade's auto-tagging, semantic neural search, or face clustering. If plain-English "find the skiing shot" search is the feature you are buying, Shade does that and JuiceMount does not, full stop. If you mostly want a fast Finder volume over storage you own, the tradeoff tilts the other way. The honest framing for the whole category is in AI search in creative storage, explained.