NSL-1: A FOUNDATIONAL RECOMMENDER
Until now, every catalogue got its own recommender, built from scratch. We found another path: one foundational model, steered by your rules rather than retrained for them.
Generic ranking works for averages. Foundational ranking works for catalogues: it learns what each person wants, and still does what the business needs.
Deployed across media, commerce, and publishing
- Streaming
- Marketplace
- Publishing
- Retail media
- Music
- Games
- Classifieds
- News
- Podcasts
- Grocery
- Fashion
- Learning
Meet NSL-1
The foundational recommender
One model, pre-trained on ranking itself, then pointed at your catalogue. It arrives already knowing how discovery behaves — so the work left to you is describing the business, not the maths.
- 01
Traceable ranking
Full visibility into how a result got where it is, with a white-box view of which signals moved it and by how much.
- 02
Signal-native
Take in any number of signals — clicks, dwell, watch time, purchases, returns, skips, explicit ratings — from any stack. For any catalogue. With any schema.
- 03
Rule-based controllability
With the capacity to adhere to conditional rules, boosts, buries, and merchandising policy, NSL-1 offers complete steerability — enabling unparalleled control over what surfaces and what does not.
- 04
Grounded results
Finally, a recommender that answers every request with the necessary context, availability, and eligibility. Never relying only on what it learned in training, and always returning items you can actually serve.
- 05
Continuous fine-tuning
NSL-1 learns and improves with every interaction, folding feedback and fresh behaviour into the next ranking automatically.
Ranking that works on behalf of catalogues
Can your recommender do this?
Catalogues of any size can fine-tune NSL-1.
Cold start
Useful ranking on day one, before a single event has been logged.
Multiple signals
Clicks, dwell, purchase, and return weighed together, not one proxy metric.
Guardrails
Business rules, boosts, buries, and exclusions applied at request time.
Groundedness
Every result checked against live availability and eligibility.
Live integrations
Catalogue and inventory read in the moment, not from last night's export.
Adaptive relevance, in the open
Most engines guess.
NSL-1 ranks.
This sandbox is a real feed. A cursor hovers, adds to cart, and favourites items, emitting weak and strong signals that re-rank the grid as they land. Switch between two users to see the same catalogue resolve to two different front pages.
NeuronSearchLab ships an MCP server, so agents can rank too
Pipelines, rules, experiments, and ranked results, exposed as tools. Point an agent at the catalogue and let it do the merchandising.
Read the docsRecommender Performance Index · v3.1
Put your catalogue to work
23 systems, scored against a frozen field on the same datasets and the same budget. Every number below is reproducible, and the ones we did not measure ourselves say so.
- Rank 1Multi-stageReproduction
ItemKNN → Multi-task ranker
Multi-task ranking over several engagement signals
Index68.8 - Rank 2Multi-stageReference
ItemKNN → MLP ranker
Classical retrieval with a neural pointwise ranker
Index66.9 - Rank 3NeighbourhoodReference
VS-KNN
Vector-multiplication session-based kNN
Index65.2 - Rank 4Linear autoencoderReference
EASEᴿ
Embarrassingly Shallow AutoEncoder
Index64.5 - Rank 5SequentialReproduction
SASRec
Self-attentive sequential recommendation
Index62.9 - Rank 6Multi-stageReference
VS-KNN → GBDT
Session-based neighbourhood retrieval with a boosted ranker
Index60.8
From the blog
Field notes
AI Agents
OpenAI MCP Marketplace: NeuronSearchLab Is Live
Jul 24, 2026
Product Strategy
Personalized Product Discovery Without Building the Full Stack
Jul 16, 2026
AI Agents
AI Agents for Recommendation Operations
Jul 7, 2026
Engineering
The Real-Time Personalization Challenge: Why Offline Metrics Don't Predict Online Success
Jun 4, 2026
Engineering
How to Control Recommendation Results with NeuronSearchLab Rules
May 21, 2026
Running at scale across media, commerce, and publishing
Index frozen 23 August 2026 · 23 systems measured