Get the most relevant search results

TopK platform combines AI-powered query understanding with adaptive ranking to provide the most relevant results in your domain.

Get Early Access

Unified search

Text, vector and metadata search in a single API call. Powers any type of retrieval.

Storing documents

Writing data is as simple as upserting a dict to a collection.

Managing collections

Documents are organized in schema-less collections by default. You can opt-in to strict types when creating a collection.

from topk import Client, Query, field, fn

client = Client(api_key="your_api_key")

# Query data from "books" collection
results = client.query(
    "books",
    select(
        # Vector search (compute similarity against a vector query)
        similarity_score=fn.vector_distance("summary_embedding", [0.5, 1.3, 5.2, ...])
        # Text search (score documents using BM25)
        text_score=fn.bm25_score()
    )
    # Full-text search (against all text-indexed fields)
    .filter(matches("harry potter"))
    # Filter by metadata
    .filter(field("published_year") >= 1996)
    # Select fields to return
    .select(
        title=field("title"),
        year=field("published_year")
    )
    # Collect TopK results by both text and vector scores
    .topk(0.8 * field("similarity_score") + 0.2 * field("text_score"), limit=3, asc=True)
)

for book in results:
    print(book["title"])

Unified search

Text, vector and metadata search in a single API call. Powers any type of retrieval.

Storing documents

Writing data is as simple as upserting a dict to a collection.

Managing collections

Documents are organized in schema-less collections by default. You can opt-in to strict types when creating a collection.

from topk import Client, Query, field, fn

client = Client(api_key="your_api_key")

# Query data from "books" collection
results = client.query(
    "books",
    select(
        # Vector search (compute similarity against a vector query)
        similarity_score=fn.vector_distance("summary_embedding", [0.5, 1.3, 5.2, ...])
        # Text search (score documents using BM25)
        text_score=fn.bm25_score()
    )
    # Full-text search (against all text-indexed fields)
    .filter(matches("harry potter"))
    # Filter by metadata
    .filter(field("published_year") >= 1996)
    # Select fields to return
    .select(
        title=field("title"),
        year=field("published_year")
    )
    # Collect TopK results by both text and vector scores
    .topk(0.8 * field("similarity_score") + 0.2 * field("text_score"), limit=3, asc=True)
)

for book in results:
    print(book["title"])

Everything you need

Without Complexity

Adaptive

Our ranking algorithms adapt to your domain through feedback to improve the relevance of results over time.

Flexible

Search multi-modal documents with powerful filtering to get the exact results your users need.

Easy to implement

TopK platform comes with batteries included. No need to integrate multiple systems to power your search.

Secure

TopK keeps data encrypted at rest and separate from other tenants. Your data is safe with us.

Low-cost

Leverage our cloud-native storage layer to get 10-100x lower cost with usage-based pricing.

Scalable

Built from first principles on object storage, our platform scales to billions of documents effortlessly.

Use Cases

RAG

Seamlessly integrate hybrid retrieval capabilities to enhance generative AI models with real-time, accurate, and relevant data.

Semantic search

Power your search solution with advanced AI query understanding understanding to deliver precise, meaningful results from unstructured datasets.

Multi-modal search

Leverage TopK to unify and optimize searches across diverse data formats like text, images, and metadata, creating a holistic retrieval experience.

Recommendations

Enhance user engagement intelligent recommendation engine, dynamically tailored to your unique data and user behavior.

Be the first one to know when we launch.

Follow our progress and get access to the first production-ready version as soon as it gets out.

Be the first one to know when we launch.

Follow our progress and get access to the first production-ready version as soon as it gets out.

Be the first one to know when we launch.

Follow our progress and get access to the first production-ready version as soon as it gets out.

Founding team

Marek Galovic

Co-Founder

Jergus Lejko

Co-Founder

Marek Dlugos

Go-To-Market