详细内容
Customers
How BrandAlley Achieved 77% Higher Conversion Rates \
See how they did it
Climatebase Increases Job Applications by 50% \
See how they did it
Partners
Education & Events
Natural Language Search on Semi-Structured Data
with Jason Liu, Staff MLE, ex-StitchFix
Cheat at Search with LLMs
with Doug Turnbull, Principal Eng, ex-Reddit
Optimize Structured Data Retrieval With Evals
with Hamel Husain, Staff MLE & Educator
From BM25 to Mixture of Encoders
at Haystack 2025
Testimonial
"We are building hyper-targeted pricing and marketing products with TBs of data & Superlinked."
Aniket Mane
VP Data at ThredUp
"We search through millions of hotels, reviews & behavioral signals with Superlinked."
James Callaghan
Head of Search at Trivago
"We match 100,000 Jira issues with multi-modal attachments to root causes with Superlinked."
Juraj Kabzan
VP Eng of Skydio
No.1 on Semi-structured Retrieval Benchmark
NDCG @10
68,78%
Superlinked (Mixture of Encoders)
Description
Uses Mixture of Encoders with Qwen3-0.6B for product description and category encoding, numerical encoders applied to product ratings, rating counts and prices. Can also generate query-specific filter predicates against materials, colors and style properties. Configured with GPT-4o for the query understanding module.
No re-ranking or metadata boosting.
61,67%
Azure AI Search (with Semantic Ranker)
Description
Azure AI Search with Semantic Re-ranker and the built-in query understanding functionality powered by OpenAI LLM API.
We implemented multiple configurations available in the Azure AI ecosystem and we took the best results provided by each - for details see link below.
57,13%
Vertex AI Search (Hybrid & Rerank)
Description
Vertex AI Search with built-in Hybrid Search configured with gemini-embedding-001 dense embedding model, splade-v3 sparse embedding model and Vertex AI RankingAPI.
The Superlinked configuration performs significantly better despite gemini-embedding-001 outperforming Qwen3-0.6B by 3 percentage points on MTEB.
51,96%
Vertex AI Discovery Engine
Description
Use SearchRequest from the Vertex AI Discovery Engine with Boost, QueryExpansion and built-in re-ranking. Tested with and without query parsing using built-in Gemini feature and took the best results.
34,75%
State of the art text embedding model
Description
The indexed JSON objects were "stringified" and embedded with Qwen3-0.6. The same model is used to encode the queries. The single dense query vector is used to retrieve the relevant results, without re-ranking.
NDCG @10\
\
68,78%\
\
Superlinked (Mixture of Encoders) 61,67%\
\
Azure AI Search (with Semantic Ranker)\
\
57,13%\
\
Vertex AI Search (Hybrid & Rerank)\
\
51,96%\
\
Vertex AI Discovery Engine\
\
34,75%\
\
State of the art text embedding model
Description
Uses Mixture of Encoders with Qwen3-0.6B for product description and category encoding, numerical encoders applied to product ratings, rating counts and prices. Can also generate query-specific filter predicates against materials, colors and style properties. Configured with GPT-4o for the query understanding module.
No re-ranking or metadata boosting.
Description
Azure AI Search with Semantic Re-ranker and the built-in query understanding functionality powered by OpenAI LLM API.
We implemented multiple configurations available in the Azure AI ecosystem and we took the best results provided by each - for details see link below.
Description
Vertex AI Search with built-in Hybrid Search configured with gemini-embedding-001 dense embedding model, splade-v3 sparse embedding model and Vertex AI RankingAPI.
The Superlinked configuration performs significantly better despite gemini-embedding-001 outperforming Qwen3-0.6B by 3 percentage points on MTEB.
Description
Use SearchRequest from the Vertex AI Discovery Engine with Boost, QueryExpansion and built-in re-ranking. Tested with and without query parsing using built-in Gemini feature and took the best results.
Description
The indexed JSON objects were "stringified" and embedded with Qwen3-0.6. The same model is used to encode the queries. The single dense query vector is used to retrieve the relevant results, without re-ranking.
Why did we create a new Information Retrieval benchmark?
This dataset is an augmented version of the WANDS (Wayfair Attribute Navigation Dataset for Search) dataset used for evaluating quality in e-commerce product search. Approximately 10,000 e-commerce product items without images (for now) including complex queries, negations, simple queries, and synthetic queries.
Build with Forward Deployed Engineers
Day 1
Scoping & intro call
to see if you have an AI Search & Matching use-case worth $1M+ of ROI
Day 14
We ship a POC with you
with an evaluation dataset we build together and an interactive UI/API
Day 28
First production A/B test
with Superlinked Cloud running in your own or our managed infrastructure
Day 42
AI in production
observable infrastructure & explainable results that fit your business objectives
Day 48
AI expansion
get into launch-measure-iterate cadence and stay at the innovation frontier of your industry
.. and onwards!
Use cases
Drive impact across your organization with Superlinked
Product\
\
Conversational search\
\
Real-time recommendations\
\
Styling assistanst Marketing\
\
Personalized emails\
\
Behavioral lookalikes\
\
Product clusters for ads Ops\
\
Support request routing\
\
Product re-duplication\
\
Missing product properties Merchandizing\
\
Purchase trend intelligence\
\
Product bundling\
\
Personalized pricing
Beyond multi-modal
Represent everything you know about your users, documents, products or jira issues with unified "omni modal" embeddings for maximum real-world retrieval relevance & control.
The AI Search Stack
Use our open source framework and server to build more reliable AI systems and onboard to Superlinked Cloud once you are ready to scale to TBs of data & millions of queries.
Experiment\
\
Describe and load your data, use Spaces to construct your indices and run queries - all in-memory within a python notebook. Deploy\
\
Deploy this code on your computer or a cloud VM, connect it to your Vector DB. Release new code by pushing it to your bucket or overwriting the local .py file. Use\
\
Push each of your entities to the auto-generated ingestion API and execute queries with query text and weight parameters using the auto-generated query API.
Resources
Learn how to build vector-powered software with VectorHub
Build RAG using LangChain & Superlinked \
Learn to build advanced RAG systems with LangChain & Superlinked. Master multi-space vector indexing, custom...\
\
#RAG\
\
#Superlinked 
How We Approached Building a Custom Steam Games Retriever with Superlinked and LlamaIndex \
Learn how to build powerful custom retrievers for RAG systems by combining Superlinked with LlamaIndex...\
\
#Vector Search\
\
#Superlinked 
Beyond Multimodal Vectors: Hotel Search With Superlinked and Qdrant \
Learn how Superlinked and Qdrant enable advanced multimodal search with natural language queries. Handle complex...\
\
#Superlinked 
Vector Search Over Complex Semi Structured Data \
\
Play\
\
#Talk 
Metadata-aware Vector Embedding MoE Models \
\
Play\
\
#Summary 
Natural Language Search on Semi-Structured Data \
\
Play\
\
#Summary
Funded by
Let's launch AI Search and Matching to production together.
Subscribe to stay updated
reCAPTCHA
Recaptcha requires verification.
I'm not a robot
reCAPTCHA
You are agreeing to our Terms and Conditions by Subscribing.
Thank you!
Your submission has been received!
Oops! Something went wrong while submitting the form.
Copyright © 2025 Superlinked Inc. All rights reserved.
Product
Use Cases How it Works VectorHub Vector DB Comparison Glossary
About
Support
Terms of Use Privacy Policy Cookie Policy Trust Center 
Links
reCAPTCHA