Astra DB Hybrid Search

Astra DB Hybrid Search is a search capability within DataStax Astra DB Serverless that combines vector and lexical search to improve retrieval accuracy for generative AI applications, particularly retrieval-augmented generation (RAG) systems.

Reviewed by 7wData

On this page

Publisher review

Astra DB Hybrid Search is a search capability within DataStax Astra DB Serverless that combines vector and lexical search to improve retrieval accuracy for generative AI applications, particularly retrieval-augmented generation (RAG) systems. It is designed for developers building AI-driven search and recommendation experiences who need higher relevance than standalone vector or keyword search can provide. The feature is currently in public preview and is available for collections in databases hosted in the AWS us-east-2 region, targeting users who prioritize search quality and are willing to work within regional and preview constraints.

Hybrid Search works by executing both a vector search and a BM25 keyword search on the same dataset, then passing the combined results through a reranking model. The reranking model used is the NVIDIA llama-3.2-nv-rerankqa-1b-v2, part of the NVIDIA NeMo Retriever text reranking microservices, which scores each result based on the original query and document content. This process can improve search relevance by up to 45% compared to vector search alone, as measured by DataStax. Developers can implement hybrid search in Python applications using version 2 of the astrapy client library, with code examples available in the official documentation and community posts.

Astra DB Hybrid Search is part of the DataStax GenAI stack, which was acquired by IBM in 2025 to complement the watsonx portfolio. It competes with other hybrid search solutions like MongoDB Atlas Search, which combines vector and full-text search, and Pinecone's serverless vector database with metadata filtering. However, Astra DB's integration with NVIDIA's reranking microservices and its foundation on Apache Cassandra for elastic scalability and near-zero latency provide a differentiator for mission-critical workloads. IBM's ownership may also appeal to enterprises already using watsonx or IBM Cloud.

A key trade-off is that hybrid search is only available in the AWS us-east-2 region, limiting deployment flexibility for global or multi-region applications. The feature is in public preview, meaning it is subject to change and not yet production-ready for all use cases. Reranking adds latency and cost, as it requires calling an external NVIDIA microservice for each query. Additionally, hybrid search requires creating a collection with specific settings that are not yet configurable through the Astra DB dashboard, forcing developers to use the Python client or API for setup.

Get the AI & data signal, daily.

335k+ subscribers read this every morning. One email, both newsletters. Unsubscribe anytime.

How it works

  1. Combines vector and BM25 search

    Performs both vector similarity and BM25 keyword searches on the same dataset to capture contextual and exact-match results.

  2. NVIDIA NeMo Retriever reranking

    Uses the NVIDIA llama-3.2-nv-rerankqa-1b-v2 cross-encoder model to rescore combined search results for higher relevance.

  3. Improves RAG relevance by 45%

    Enhances retrieval-augmented generation systems by returning more accurate context, reducing LLM hallucination risk.

  4. Python client support

    Available via the astrapy Python client version 2, with code examples for creating collections and performing hybrid queries.

  5. Metadata filter integration

    Applies optional metadata filters before vector and lexical searches to narrow the dataset for targeted retrieval.

  6. Public preview on AWS us-east-2

    Currently limited to databases in the AWS us-east-2 region, with ongoing development and potential feature changes.

  7. Part of DataStax GenAI stack

    Integrated with Astra DB Serverless and IBM watsonx, supporting enterprise AI workloads with elastic scalability.

Strengths and trade-offs

Strengths

  • Improves search relevance by up to 45% compared to vector search alone, as demonstrated in DataStax benchmarks.
  • Combines vector search and BM25 keyword search with NVIDIA NeMo Retriever reranking for more accurate results.
  • Supports Python integration via the astrapy client library version 2, with detailed code examples in official documentation.
  • Built on Apache Cassandra for elastic scalability and near-zero latency, suitable for mission-critical AI workloads.

Trade-offs

  • Hybrid search is only available for collections in databases in the AWS us-east-2 region, limiting geographic deployment options.
  • The feature is in public preview and subject to change, with no production-level stability guarantees yet.
  • Reranking requires calls to external NVIDIA microservices, adding latency and potential cost per query.
  • Collection setup for hybrid search requires using the Python client or API, as dashboard configuration is not yet supported.

Pricing context

Pricing is not publicly detailed in the provided sources; Astra DB Serverless typically offers a free tier with pay-as-you-go for additional usage, but hybrid search-specific costs (e.g., NVIDIA reranking microservice fees) are not specified.

Getting started with Astra DB Hybrid Search

  1. Sign up for Astra DB

    Create a DataStax Astra DB account at the official website. Choose the Serverless plan and select the AWS us-east-2 region for your database, as hybrid search is currently limited to this region.

  2. Install the astrapy client

    Install version 2 of the astrapy Python client library using pip. Run 'pip install astrapy>=2.0.0' in your terminal to ensure you have the latest version that supports hybrid search features.

  3. Connect to your database

    Use your Astra DB application token and database ID to establish a connection. In your Python script, create an AstraDB client instance with these credentials to authenticate and access your collections.

  4. Create a hybrid search collection

    Define a collection with hybrid search settings using the astrapy client. Specify the vector dimension, metric, and enable hybrid search by setting the 'hybrid' parameter to True. This cannot be done through the dashboard.

  5. Run a hybrid search query

    Insert documents into your collection, then execute a hybrid search by calling the 'find' method with a query string and optional metadata filters. The client automatically combines vector and BM25 searches with NVIDIA reranking.

Frequently Asked Questions

What is Astra DB Hybrid Search and how does it work?

Astra DB Hybrid Search is a DataStax feature that combines vector similarity and BM25 keyword searches on the same dataset. Results are then reranked using NVIDIA's llama-3.2-nv-rerankqa-1b-v2 model to improve relevance for generative AI applications like RAG.

How does Astra DB Hybrid Search improve RAG accuracy?

By executing both vector and BM25 searches, then reranking combined results with NVIDIA's cross-encoder model, hybrid search can improve retrieval relevance by up to 45% compared to vector search alone. This reduces hallucination risk in LLM outputs by providing more accurate context.

What are the limitations of Astra DB Hybrid Search during public preview?

Hybrid search is only available for collections in the AWS us-east-2 region. It is in public preview, meaning features may change without notice. Setup requires the Python client or API since dashboard configuration is not yet supported. Reranking adds latency and potential costs.

How do I implement Astra DB Hybrid Search in Python?

Use version 2 of the astrapy Python client library. Create a collection with specific settings for hybrid search, then perform queries that combine vector and lexical searches. Official documentation and community posts provide code examples for setup and querying.

How does Astra DB Hybrid Search compare to MongoDB Atlas Search?

Both combine vector and full-text search, but Astra DB integrates NVIDIA NeMo Retriever reranking for higher relevance. Built on Apache Cassandra, it offers elastic scalability and near-zero latency. IBM's ownership may appeal to enterprises using watsonx or IBM Cloud.

What costs are associated with Astra DB Hybrid Search?

Pricing for hybrid search itself is not publicly detailed. Astra DB Serverless typically offers a free tier with pay-as-you-go for additional usage. However, reranking requires calls to external NVIDIA microservices, which may incur separate fees not specified in available sources.

Alternatives

How Astra DB Hybrid Search compares

Direct head-to-head against 3 competitors. Picked by 7wData.

This tool

Astra DB Hybrid Search

Pricing
Pricing is not publicly detailed in the provided sources; Astra DB Serverless typically offers a free tier with pay-as-you-go for additional usage, but hybrid search-specific costs (e.g., NVIDIA reranking microservice fees) are not specified.
Target
Astra DB Hybrid Search is a search capability within DataStax Astra DB Serverless that combines vector and lexical search to improve retrieval accuracy for generative
Strength
Improves search relevance by up to 45% compared to vector search alone, as demonstrated in DataStax benchmarks.
Watch for
Hybrid search is only available for collections in databases in the AWS us-east-2 region, limiting geographic deployment options.

Weaviate

Pricing
Free tier (sandbox), Enterprise: Custom/Contact sales
Target
Hybrid search (vector + BM25 + metadata filters) at scale
Deployment
Self-hosted or managed cloud
Strength
Native hybrid search with built-in BM25 and reranking modules
Watch for
Complex self-hosted setup; pricing escalates with cluster size

Pinecone

Pricing
Free tier (1 pod), Standard: $70/month per pod, Enterprise: Custom
Target
Zero-ops managed vector search at any scale
Deployment
Fully managed cloud only
Strength
Simplest managed experience with no infrastructure overhead
Watch for
No native hybrid search; requires separate keyword index

Qdrant

Pricing
Free tier (1GB), Cloud: from $25/month, Enterprise: Custom
Target
Budget-conscious teams under 50M vectors
Deployment
Self-hosted or managed cloud
Strength
Best free tier with generous 1GB storage and no time limit
Watch for
Hybrid search is experimental; BM25 support not production-ready

User reviews

No user reviews yet. Be the first to write one.

Sources

Reporting on this tool draws on these publicly available sources.

  1. www.linkedin.com
  2. www.ibm.com
  3. dev.to
  4. docs.datastax.com
  5. www.businesswire.com
  6. www.datastax.com