Weaviate Vector Store

The Weaviate Vector Store node enables integration with a Weaviate vector database, either hosted on the cloud or deployed locally. This node allows storage and retrieval of vectorized documents for downstream LLM tasks. 

Inputs

  • Documents – Vectorized document embeddings to store
  • Questions – Incoming queries to search similar vectors

Outputs

  • Documents – Retrieved documents matching the query
  • Answers – Search results (most relevant document chunks)
  • Questions – The original question passed along

Configuration

GUI

  • Type of Weaviate host
    • Weaviate cloud server
    • Your own Weaviate server
  • Host
    • Cloud – your-instance-name.weaviate.cloud
    • Local – typically localhost
  • Port
    • Default – 8080
    • Required when “Your own Weaviate server” is selected
  • gRPC Port (only for local setup)
    • Default – 50051
  • API Key – Enter your Weaviate API key (if applicable for authentication)
  • Retrieval Score – Define the threshold for similarity
    • Options – Related, Very Related
    • This determines what results are returned from the vector store.
  • Collection – The name of the collection or index to read/write vectors
    • Example – aparavi
    • Must be lowercase with optional hyphens
  • Example
    • Host: localhost
    • Port: 8080
    • gRPC Port: 50051
    • Retrieval Score: Related
    • Collection: aparavi

Weaviate Vector Store supports several deployment modes:

Local Mode

This guide walks you through configuring Ngrok on your local machine

Cloud Mode

Connects to a Weaviate cloud instance.

{
"url": "https://your-cluster-id.weaviate.cloud",
"api_key": "your-api-key",
"class_name": "Document"
}

Embedded Mode

Uses an embedded Weaviate instance within Aparavi.

{
  "embedded": true,
  "persistence_path": "./weaviate-data",
  "class_name": "Document"
}

Advanced Settings

  • Batch Size – Number of objects to batch insert
    • Default – 100
    • Notes – Higher values improve write performance
  • Vector Index Type – Type of vector index
    • Default – “hnsw”
    • Options – hnsw, flat
  • Vector Distance – Distance metric for similarity
    • Default – “cosine”
    • Options – cosine, dot, l2-squared
  • Auto Schema – Automatically generate schema
    • Default – true
    • Notes – Set to false for custom schemas
  • Tenant – Multi-tenancy identifier
    • Default – null
    • Notes – For multi-tenant deployments

Example Usage

Basic RAG Pipeline

This example shows how to use Weaviate Vector Store in a basic Retrieval Augmented Generation (RAG) pipeline:

  1. Connect a Document Parser to extract text from documents
  2. Connect a Preprocessor to clean and prepare the text
  3. Connect an Embeddings node to convert text to vector embeddings
  4. Connect the Embeddings output to the Weaviate Vector Store Documents input
  5. Connect a question input to the Weaviate Vector Store Questions input
  6. Connect the Weaviate Answers output to an LLM for generating responses

Best Practices

  • Use separate classes for different data domains
  • Define custom schemas for better control over property indexing
  • Use batch imports for large datasets to improve performance
  • Adjust HNSW index parameters based on your precision vs. speed requirements
  • Use GraphQL filtering capabilities for complex queries
  • When using a local server, make sure the Weaviate instance is running and accessible at the specified host and port.
  • Use matching embedding dimensions and model types between your embedding node and Weaviate collection schema.

 

Troubleshooting

Connection Problems

  • Connection refused – Verify host and port settings
  • Authentication failure – Check API key validity
  • Timeout errors – Check network connectivity and Weaviate server status

Query Performance

  • Slow queries – Optimize index parameters (ef, maxConnections)
  • Memory errors – Check Weaviate server resources
  • Poor search results – Ensure vector dimensions match between embeddings and schema

Technical Reference

For detailed technical information, refer to: