Skip to content

Vector Search Building Block

Build high-performance hybrid search solutions — combining semantic vector search with BM25 keyword search — using IBM watsonx.data OpenSearch and IBM watsonx.ai embeddings. Ingest documents from IBM Cloud Object Storage, parse them with IBM Docling, generate dense embeddings, and run hybrid search queries that outperform pure vector-only search.

GitHub Repository

The complete source code and examples are available in the GitHub repository:

Building Blocks - Vector Search


Overview

The Vector Search building block provides standalone ingestion and hybrid search services for GenAI pipelines. Each integration ships a complete FastAPI asset, a Bob Mode, and a Bob Skill — ready to deploy on IBM Cloud, Red Hat OpenShift, or locally.

Recommendation: Use hybrid search (vector + BM25) rather than pure vector-only search. Combining dense semantic search with BM25 keyword matching consistently produces superior retrieval accuracy for RAG applications.


When to Use

Scenario Building Block
Build a standalone search service for documents stored in IBM COS OpenSearch
Best retrieval accuracy with hybrid search (vector + BM25) OpenSearch — recommended
Power the retrieval layer for a RAG pipeline OpenSearch
Need IBM HCD (Astra DB) serverless vector collections DataStax Astra DB

Supported Integrations

Available Integrations

  • OpenSearch — IBM watsonx.data managed OpenSearch with k-NN HNSW + BM25 hybrid search ✅ Available Now
  • DataStax Astra DB — IBM HCD serverless Cassandra-backed vector storage ✅ Available Now

Location: opensearch/ IBM Products: IBM watsonx.data (OpenSearch), IBM watsonx.ai, IBM COS

Hybrid search engine on IBM watsonx.data managed OpenSearch — combines k-NN vector search with BM25 keyword search. Uses HNSW indexes via the k-NN plugin. This is the recommended backend for all new hybrid search and RAG projects.

Key Features: - OpenSearch data ingestion FastAPI service - Hybrid search: dense k-NN vector search + BM25 keyword search with score fusion - IBM watsonx.ai embedding integration (ibm/slate-125m-english-rtrvr, dim=768) - Bob Mode: opensearch-builder.zip - Bob Skill: opensearch-vector-search.zip

View OpenSearch Details →


DataStax Astra DB

Location: datastax-astradb/ IBM Products: IBM HCD (DataStax Astra DB), IBM watsonx.ai, IBM COS

Vector search using DataStax Astra DB — part of the IBM Cloud HCD (Hyper-Converged Database) portfolio. Ingest documents from IBM COS, generate dense embeddings with IBM watsonx.ai, and perform ANN cosine similarity search.

Key Features: - Astra DB vector ingestion FastAPI service - ANN cosine similarity search via astrapy Data API - Globally distributed, serverless Cassandra-backed storage - Bob Mode: astradb-vector-builder.zip - Bob Skill: astradb-vector-setup.zip

View DataStax Astra DB Details →


Search Strategy Comparison

Approach How It Works Results
Hybrid Search (recommended) Dense vector + BM25 keyword with score fusion ✅ Best accuracy — catches both semantic and exact matches
Vector-only search Dense vector similarity (cosine / ANN) ⚠️ Misses keyword-specific queries
Keyword-only (BM25) Term frequency / inverse document frequency ⚠️ Misses paraphrase and semantic matches

OpenSearch vs Astra DB

OpenSearch Astra DB
Search type Hybrid (vector + BM25) Vector ANN (cosine)
IBM product IBM watsonx.data IBM HCD
Architecture Managed cluster Serverless
Best for RAG + hybrid search Serverless global vector storage
Recommendation ✅ New projects When IBM HCD is required

Embedding Models

Model ID Dimension Language Use Case
ibm/slate-125m-english-rtrvr 768 English Recommended for English RAG
ibm/slate-30m-english-rtrvr 384 English Lightweight English RAG
intfloat/multilingual-e5-large 1024 Multi Multilingual RAG

Prerequisites

  • IBM Cloud API keycreate at IBM Cloud IAM
  • Python 3.10+
  • IBM watsonx.ai project — note Project ID and instance URL
  • IBM Cloud Object Storage bucket with source documents
  • IBM watsonx.data OpenSearch instance (for OpenSearch) or Astra DB instance (for DataStax)

Use Cases

  • Semantic Search — Find documents based on meaning, not just keywords
  • RAG Pipelines — Retrieval-augmented generation for LLMs
  • Knowledge Bases — Build searchable knowledge repositories
  • Document Discovery — Find similar documents across large collections
  • Question Answering — Retrieve relevant context for Q&A systems

Resources


Support

For issues or questions, please refer to the GitHub repository or open an issue.