- 01rank candidates by intent
- 02weight by behavioural signal
- 03diversify the top-k
- 04serve top-k
Search and recommendations that work on request one. The same model reads your text, looks at your images, and ranks the catalogue — in 60 ms, on the 86% of sessions with no history at all.
RUS_ID:SOLENYA|
Fig_1.0: the_source
Jaguar Enabled
v.3.1.4Built on the shoulders of giants
MilvusPostgreSQLGraphQLNixFastAPIGCPHelmMCPPrefectTemporalKafkaBunClickHouseNext.jsPyTorchMilvusPostgreSQLGraphQLNixFastAPIGCPHelmMCPPrefectTemporalKafkaBunClickHouseNext.jsPyTorch
One ranker, every surface
Recommendations - but everywhere.
The carousel, the listing, the search bar, the agent. One model ranks all of it - no second vendor, no second pipeline.
Fig 4.0: recommendations_everywhereSame ranker
The status quo
Recommended for you · Top 3
#1
Brine Trail boot£128
#2
Salt Pier hiker£139
#3
Pickle Ridge boot£146
k = 3 · the rest of the catalogue is invisible here
SolenyaScroll ↓
Listing · /boots
All✓ Boots✓ Under £150WaterproofIn stock
Brine Trail boot£128
Salt Pier hiker£139
Pickle Ridge boot£146
Dill Marsh chukka£124
Caraway Crag£152
Fennel Field£118
Mustard Mile£134
Tarragon Trek£141
Saffron Slope£129
Coriander Cove£148
Juniper Junction£137
Rosemary Reef£155
Same ranker, same model, every surface — no second vendor for "recommendations," no frozen list waiting for next week's re-train.
Zero-shot · per request · per user
Every click reshapes the next page. Yours, not the cohort's.
Global sorts hand every shopper the same five top picks - and they're the first to sell out. Solenya re-ranks per request, so what surfaces first is what's in *your* size, in stock, right now.
Fig 5.0: live_rerank · per_requestSame catalogue
Everyone else
Global sort · /boots
#1
Brine Trail bootOut of stock
#2
Salt Pier hikerOut of stock
#3
Pickle Ridge boot1 left
#4
Dill Marsh chukka£124
#5
Caraway Crag£152
#6
Fennel Field£118
top picks sell out first · same list for everyone
Solenya
Re-rank · /boots · 38 ms
Brine Trail boot£128
Salt Pier hiker£139
Pickle Ridge boot£146
Dill Marsh chukka£124
Caraway Crag£152
Fennel Field£118
re-rank #N · per request · per anonymous session
Same catalogue, same second - different ranking. No segment, no cookie, no nightly retrain. One click reshapes the next page; the anonymous visitor next to you gets a different one.
The day-zero gap
Legacy never catches up. Every new SKU resets the clock.
New stock, new brands, new categories - each one knocks a click-history ranker back down. We rank from request one and don’t flinch when the catalogue moves.
Fig 2.0: cold_start_curve
SolenyaLegacy
Click-history rankers warm up over weeks — then every new SKU, brand, or category resets the clock. Solenya starts at par on request one and stays there.
More than a wrapper
Multimodal discovery, designed for multimodal agents.
Seamless hand-off. Agents - but personalised. Eval-led at every step.
Fig 3.0: agent_calls_solenyaMCP
user
find me a warm waterproof boot under £150
agent · tool_call
solenya.search()POST /mcp
{
"query": "warm waterproof boot under £150",
"modality": [
"text",
"image"
],
"top_k": 4
}response · 4 results · 58 ms
Brine Trail boot£128
Salt Pier hiker£139
Pickle Ridge boot£146
Dill Marsh chukka£124
Why we can do this
From separate jars to one kitchen.
Discovery has been stuck in separate jars - text in one vendor, images in another, behaviour in a third, products in a silo. We unified them to unlock a new generation of e-commerce experiences.
Text
Image
Behaviour
Products
One model1024-d
Same space. One ranker. Sub-60 ms P50.
Fig 1.0: unified_embedding_space
What that unlocks
Search by photo. Rank by intent. Recommend by both.
More than a thousand words. One model that reads, sees, and remembers the catalogue means the same query works in every modality - tap a pill to see it on a real listing.
Image
Product
Text
User
Solenya
Composed Image
Pair a reference image with a text edit — “like this, but in red” — and retrieve real SKUs from inventory.
Fig 9.5: modality_playgroundReverse Image
Listing · /boots · 60 ms
textimageproductuser
Brine Trail boot£128
Salt Pier hiker£139
Pickle Ridge boot£146
Dill Marsh chukka£124
Caraway Crag£152
Fennel Field£118
Show me what looks like this.
Mise en place · before your data arrives
Your catalogue is a teaspoon. Pretraining is the supermarket.
Internal datasets are too small. A single public benchmark isn’t much bigger. Solenya understands every new brand, trend and category. From day zero.
Fig 3.0: pretraining_mix
YoursPublicSolenya
Your catalogueone merchant
RetailRocketbenchmark
YOOCHOOSEbenchmark
Taobao User Behaviorbenchmark
Amazon Reviews 2023benchmark
Solenya midtraining mixwhat you walk in with
Each public benchmark is one ingredient. Your catalogue is another. We pretrain on the whole pantry - your shoppers’ clicks just pick the recipe.
Post-training · the staff tasting
Discovery in the Post-training era.
Pretraining stocks the pantry. Post-training plates it. Before service the kitchen tastes its own work — seasoning it to perfection. Your shoppers are diners, not test subjects.
Fig 3.1: post_training_deltas
Printed recipeRehearsed edits
- 01rank candidates by intentsize > brand at browse
- 02weight by behavioural signaldemote 'sale' for premium
- 03diversify the top-k
- 04serve top-k+ outfit pairing on PDP
same printed recipe, every kitchen — only one of them tastes its work before service
Service · freshness, not yesterday’s hits
Recommendations that stay fresh. Catalogues that don’t feel smaller over time.
Freshness isn’t a campaign — it’s the willingness to keep exploring. Rankers that only double down on yesterday’s hits look like they’re optimising while the catalogue gets quietly smaller. Ours keeps reaching past the obvious, so the long tail, the re-emerging picks and the brand-new arrivals stay in play — and the storefront keeps feeling whole.
Fig 4.0: rut_vs_freshness
regimelegacy_loop
metric ↑ · catalogue ↓Bestseller #1served yesterday
Bestseller #1served today
Bestseller #1served tomorrow
Bestseller #1served next week
reward_estimatelooks great, isn’t
regimesolenya_fresh
tracks shifting taste
StaleFresh
sku · 01
sku · 02
sku · 03
sku · 04
sku · 05
sku · 06
tastes shift — the catalogue stays whole
Every channel · one model
Same model, every level. Not just the search bar.
Inside Listings View, traffic splits across Browse, Search, Top-k Rec and Chat. Doubling the chat bubble shifts purchase by a chat bubble's worth. Solenya lifts the whole mix. Drag the sliders - or watch the demo run itself.
Fig 3.1: funnel_lift_simulatorMix-weighted conversion
Shopper funnel · Listings View runs on four channels
Session → purchase2.10%Baseline 2.10%
Δ vs baseline+0.00%+0.0% relative
Solenya master · lift every channel+0%
- Browse(82% of listings)+0%
- Search(13% of listings)+0%
- Top-k Rec(3% of listings)+0%
- Chat(2% of listings)+0%
Drag a slider, or watch the demo.
Lifting one bubble lifts the dial by a bubble's worth. Solenya lifts every bubble — same model, every level.
Source: anonymised production telemetry; Solenya ranks every channel with the same zero-shot model.
Every request · ranked fresh
Your CDN hides 86% of your traffic. Personalising wakes it up.
Storefront origins only ever see the unique 14% - the rest gets cached at the edge. Personalise everything and that 86% has to be ranked on every hit. Roughly 7× the load your stack handles today. We absorbed the 7× - delivering a new, personalized experience to users in real-time.
Fig 3.1: traffic_mosaicTop strip = what your origin sees · The rest = what your CDN hides
- Solenya — ranks fresh, every request
- Legacy — cacheable or search-only
Same model, every request. The cache no longer hides the hard part.
Source: Solenya production telemetry. Multiplier = 1 / origin-visible share; assumes uniform per-request ranking cost.
Ship Discovery That Earns Its Numbers
Personalized results from day one: the moment a new user arrives, the moment a new product arrives, the moment a new category is launched, the moment you integrate.
Plugs into your stack
ShopifyGraphQLMCPOAuthOIDCShopifyGraphQLMCPOAuthOIDC
Blogs
Open-notebook research. Methods, confidence intervals, and the experiments that didn't work.
