Recommendation Engine for Streaming Services
Let viewers say what they want tonight, and rank your catalog against it with a written reason under every pick. The catalog, the interface and the viewer stay yours.
TasteRay is a recommendation API for streaming services. You send a viewer's words and up to ten titles from your catalog, and one endpoint returns them ranked, each with a match score and a written reason. It needs no viewing history to start, so someone who subscribed today still gets a usable answer.
Viewing history tells you what someone watched last month. It says little about tonight, and nothing at all about a person who joined this morning. TasteRay reasons from what the viewer tells you in their own words, plus any history you choose to send.
Your viewers are being taught to ask
On 3 November 2026 Netflix announced AI Discover, a test tab where a small group of US members pick a theme by mood, company or time available. On 6 August 2026 Disney described a Disney+ beta for finding something by describing it. On 28 July 2026 TechCrunch reported an experimental conversational search test at HBO Max.
All three are limited tests inside their owners' apps. They still set an expectation, and a smaller service has no research group to answer it with. Recommendations have carried this weight for a long time: in 2015, Netflix researchers wrote that its recommender system influenced choice for about 80% of hours streamed.
One tired Sunday, ten titles from a catalog
We sent one sentence a viewer might say, three preferences, and ten titles that included Grown Ups 2, The Raid and Midsommar. The top three came back, each with a reason a viewer can read and disagree with.
{
"vertical": "movies",
"context": {
"profile": "Niche SVOD subscriber. Sunday night, wrung out after a long week, wants something gentle and good. Nothing dumb.",
"preferences": ["quiet", "humane", "well made"]
},
"items": [
{"id": "m1", "name": "Paterson", "description": "A bus driver in New Jersey writes poems in a notebook through an ordinary, gentle week.", "metadata": {"year": 2016, "runtime_minutes": 118}},
{"id": "m2", "name": "Lost in Translation", "description": "Two lonely strangers in Tokyo form a quiet bond over a few sleepless nights.", "metadata": {"year": 2003, "runtime_minutes": 102}},
// ...and 8 more titles: Grown Ups 2, The Raid, Midsommar and others
],
"options": {"count": 3}
}meta.processing_time_ms 2,099. The model's sentences are verbatim.Paterson
96 out of 100, taste_match_score
You are looking for a warm blanket after a long week, not another performance. 'Paterson' is the cinematic equivalent of a hot cup of tea and a deep breath—it is quiet, humane, and deeply well-made.
- This film perfectly mirrors your request for 'quiet' and 'humane.' It follows a bus driver through his ordinary, gentle week, celebrating the beauty in the mundane without trying to force any drama.
- At 118 minutes, it is a gentle commitment that doesn't demand high-energy engagement. It’s the perfect Sunday night companion to help you decompress and let your mind settle.
Aftersun
91 out of 100, taste_match_score
Sometimes 'gentle' isn't about escaping—it’s about connection. 'Aftersun' is a profoundly humane and beautifully crafted look at memory, offering the emotional resonance you crave from 'well-made' cinema.
- It is a masterpiece of modern indie cinema that completely respects your intelligence. There are no cheap twists or loud shocks, just a deeply felt, quiet exploration of a father-daughter bond.
Lost in Translation
85 out of 100, taste_match_score
A gentle film about two lonely people finding comfort in each other's quiet company. 'Lost in Translation' is a masterclass in mood, offering a beautiful, melancholic warmth that is the perfect antidote to a grueling week.
- At 102 minutes, it’s a manageable length for a Sunday night. It’s a film about finding a moment of peace in a chaotic world, exactly what you need to feel restored.
Fields shown: item.name, metadata.taste_match_score, explanation.why_match, explanation.key_factors.
What changes in a streaming integration
- Rights and territories
- Send only the titles a viewer can play in their territory today. With
itemsset, the API only returns items from the list you send, so an expired licence never needs an apology. - The first session
- A new subscriber has no history to mine. One sentence in
context.profileis enough for a ranked answer, and the same call takescontext.historylater, up to 50 rated items. - Several languages
options.languagesets the language of every generated string, so a regional service can show reasons in its viewers' language from one integration. Check the output in each language you ship.- The living room
- There is no TV SDK. Your server calls the API and your app renders the result. A response takes a few seconds, so TV rows should come from your own cache, filled ahead of time.
- Facts on the card
- Show runtime, cast, rating and availability from your own metadata. The reason is model text about fit: render it as an opinion, and read samples for wrong facts before you launch.
Sizing the request budget
One request ranks one list, and a cached row costs nothing to show again. If you precompute by theme and audience segment, the bill follows the size of that grid. Your number of subscribers does not enter into it until you personalize per viewer.
The table is arithmetic on the published limits, with a 30-day month. A viewer typing free text is always a live request, so add those on top.
| Example | Calculation | Requests a month | Tier that fits |
|---|---|---|---|
| Evaluation | 4 themes, refreshed by hand a few times | Under 100 | Free, 1,000 a month |
| Small service | 12 themes × 20 segments, refreshed daily | 7,200 | Basic, $49 a month for 10,000 |
| Growing service | 20 themes × 100 segments, refreshed daily | 60,000 | Pro, $199 a month for 100,000 |
| Per-viewer rows | 5 rows × 20,000 active viewers, refreshed daily | 3,000,000 | Enterprise, custom limits |
When this is the wrong tool
If you need a sub-second answer on every row of a home screen, models trained on your own interaction logs fit better than a call that generates text. The same goes for a service whose recommendations are its main edge over competitors. Build vs buy sets out how to decide, including prices from other vendors.
If you want the step-by-step version of the mood-based tab, read the implementation guide. For larger operators, see Enterprise.
TasteRay is not affiliated with Netflix, Disney or HBO Max. Their facts come from the sources linked on this page, as of the dates shown.
Sources and further reading
- Introducing AI Discover (Netflix, 3 November 2026)
- Beta Testing AI-Powered Search and Discovery on ESPN and Disney+ (The Walt Disney Company, 6 August 2026)
- HBO Max embraces vertical video with a new Shorts feed (TechCrunch, 28 July 2026)
- The Netflix Recommender System: Algorithms, Business Value, and Innovation (Gomez-Uribe and Hunt, ACM TMIS, 2015)
- TasteRay API documentation
FAQs about Recommendation Engine for Streaming Services
How much viewing data does a streaming service need to start with TasteRay?
A streaming service needs no viewing data to start with TasteRay. Each request carries its own context, so a sentence from the viewer and a list of titles is enough. You can add up to 50 rated items of history per request when you have them, and you decide what leaves your system.
Can the TasteRay API recommend only titles from our catalog?
Yes. Send your titles in the items field and the API ranks that list instead of generating its own, returning up to ten per request. Filter for rights and territory before you send it. Each result carries the `id` and the exact `name` you sent, so match on `item.id`. Results not in your list are dropped, so fewer than you asked for can come back.
Is the TasteRay API fast enough for a TV interface?
Only with a cache in front of it. Fast mode took a few seconds in most of our tests on 5 November 2026, with slower outliers above 20 seconds, and every response reports its own processing time. Precompute rows by theme and segment, then serve them from your own store.
Is TasteRay the same technology as Netflix AI Discover?
Nobody outside Netflix can say. Netflix describes a new intelligence layer but names no model or vendor, as of its 3 November 2026 post. TasteRay is a separate company and is not affiliated with Netflix. It offers a documented API that any service can call and test against its own catalog.