Vector Database Integrations

Configure vector databases for RAG (Retrieval Augmented Generation)

Overview

Open WebUI supports 9 vector database options for storing and retrieving document embeddings in RAG workflows. Choose based on your deployment requirements, scale, and features needed.

Supported Vector Databases

ChromaDB

Default, embedded, zero-config

Qdrant

High-performance, production-ready

Milvus

Scalable, cloud-native

PGVector

PostgreSQL extension

Elasticsearch

Full-text + vector search

OpenSearch

Open source Elasticsearch alternative

Pinecone

Managed cloud service

S3Vector

S3-backed vector storage

Oracle 23ai

Oracle database with vector support

ChromaDB (Default)

Embedded vector database with zero configuration.

Configuration

ChromaDB works out of the box with no configuration:

# Already included in requirements.txt
chromadb==1.4.1

File: backend/requirements.txt:55

Features

• Zero configuration
• Embedded mode (no separate server)
• Automatic persistence
• Metadata filtering
• Hybrid search

Use Cases

Single-server deployments

Development and testing

Small to medium datasets

Qdrant

High-performance vector database optimized for production.

Installation

1. Install Client

   # Already included in requirements.txt
   qdrant-client==1.16.2
   
2. Deploy Qdrant

   Docker

   docker run -p 6333:6333 \
     -v qdrant_storage:/qdrant/storage \
     qdrant/qdrant
   
3. Configure Open WebUI

   Set environment variables or configure in admin panel

Configuration

VECTOR_DB=qdrant
QDRANT_URL=http://localhost:6333
QDRANT_API_KEY=optional-api-key

Features

• High performance
• Distributed deployment
• HNSW indexing
• Filtering and payload support
• Cloud-managed option

File: backend/requirements.txt:123

Milvus

Scalable, cloud-native vector database.

Installation

1. Install Client

   # Already included in requirements.txt
   pymilvus==2.6.8
   
2. Deploy Milvus

   Docker Compose

   # Download Milvus docker-compose.yml
   wget https://github.com/milvus-io/milvus/releases/download/v2.3.0/milvus-standalone-docker-compose.yml -O docker-compose.yml
   docker-compose up -d
   

Configuration

VECTOR_DB=milvus
MILVUS_URL=http://localhost:19530
MILVUS_TOKEN=optional-token

Features

• Horizontal scalability
• Multiple index types (HNSW, IVF, etc.)
• GPU acceleration support
• Time travel queries
• Partition support

File: backend/requirements.txt:122

PGVector

PostgreSQL extension for vector similarity search.

Installation

1. Install Dependencies

   # Already included in requirements.txt
   psycopg2-binary==2.9.11
   pgvector==0.4.2
   
2. Setup PostgreSQL with pgvector

   PostgreSQL

   -- Enable pgvector extension
   CREATE EXTENSION IF NOT EXISTS vector;
   

Configuration

VECTOR_DB=pgvector
DATABASE_URL=postgresql://user:password@localhost:5432/openwebui

Features

• SQL-based queries
• ACID compliance
• Existing PostgreSQL infrastructure
• Familiar tooling
• Cost-effective

File: backend/requirements.txt:116-117

Elasticsearch

Full-text search with vector capabilities.

Installation

# Already included in requirements.txt
elasticsearch==9.3.0

Configuration

Self-Hosted

VECTOR_DB=elasticsearch
ELASTICSEARCH_URL=http://localhost:9200
ELASTICSEARCH_API_KEY=optional

Features

• Hybrid search (full-text + vector)
• Mature ecosystem
• Advanced analytics
• Scalable architecture

File: backend/requirements.txt:125

OpenSearch

Open source alternative to Elasticsearch.

Installation

# Already included in requirements.txt
opensearch-py==3.1.0

Configuration

VECTOR_DB=opensearch
OPENSEARCH_URL=http://localhost:9200
OPENSEARCH_USERNAME=admin
OPENSEARCH_PASSWORD=admin

Features

• k-NN plugin for vector search
• Apache 2.0 license
• AWS managed service available
• Compatible with Elasticsearch APIs

File: backend/requirements.txt:57

Pinecone

Managed cloud vector database.

Installation

# Already included in requirements.txt
pinecone==6.0.2

Configuration

VECTOR_DB=pinecone
PINECONE_API_KEY=your-api-key
PINECONE_ENVIRONMENT=us-west1-gcp
PINECONE_INDEX=openwebui

Features

• Fully managed
• No infrastructure management
• Auto-scaling
• Low latency
• Free tier available

File: backend/requirements.txt:126

Oracle 23ai

Oracle Database with AI Vector Search.

Installation

# Already included in requirements.txt
oracledb==3.4.2

Configuration

VECTOR_DB=oracle
ORACLE_DSN=localhost:1521/FREEPDB1
ORACLE_USER=vector_user
ORACLE_PASSWORD=password

Features

• Integrated with Oracle Database
• ACID transactions
• Enterprise features
• SQL-based vector search

File: backend/requirements.txt:127

Configuration Comparison

Database	Setup Complexity	Performance	Scalability	Cost
ChromaDB	⭐ Easy	⭐⭐⭐ Good	⭐⭐ Medium	Free
Qdrant	⭐⭐ Medium	⭐⭐⭐⭐⭐ Excellent	⭐⭐⭐⭐ High	Free/Paid
Milvus	⭐⭐⭐ Complex	⭐⭐⭐⭐⭐ Excellent	⭐⭐⭐⭐⭐ Highest	Free/Paid
PGVector	⭐⭐ Medium	⭐⭐⭐ Good	⭐⭐⭐ High	Free
Elasticsearch	⭐⭐⭐ Complex	⭐⭐⭐⭐ Very Good	⭐⭐⭐⭐ High	Free/Paid
OpenSearch	⭐⭐⭐ Complex	⭐⭐⭐⭐ Very Good	⭐⭐⭐⭐ High	Free
Pinecone	⭐ Easy	⭐⭐⭐⭐ Very Good	⭐⭐⭐⭐⭐ Highest	Paid
Oracle 23ai	⭐⭐⭐ Complex	⭐⭐⭐⭐ Very Good	⭐⭐⭐⭐ High	Paid

Switching Vector Databases

Switching vector databases will require re-indexing all documents. Existing embeddings are not automatically migrated.

1. Backup Data

   Export your documents from the current database

2. Configure New Database

   Update environment variables or admin settings

3. Restart Open WebUI

   docker restart open-webui
   
4. Re-index Documents

   Upload documents again or trigger re-indexing

RAG Configuration

Embedding Models

Open WebUI uses embedding models to convert text to vectors:

# Default sentence-transformers
sentence-transformers==5.2.2

Common models:

• all-MiniLM-L6-v2 (default, 384 dims)
• all-mpnet-base-v2 (768 dims)
• text-embedding-3-small (OpenAI, 1536 dims)

File: backend/requirements.txt:60

Chunk Configuration

Control how documents are split:

CHUNK_SIZE=1000
CHUNK_OVERLAP=200

Retrieval Settings

TOP_K=5  # Number of chunks to retrieve
SIMILARITY_THRESHOLD=0.7  # Minimum similarity score

Performance Tuning

Index Type

Choose appropriate index (HNSW, IVF, FLAT) based on dataset size

Dimension Reduction

Use smaller embedding models for better performance

Batch Processing

Index documents in batches for efficiency

Caching

Enable result caching for repeated queries

Troubleshooting

Connection Failed

1. Verify database is running: docker ps or check service status
2. Check connection string/URL format
3. Verify credentials (API key, username, password)
4. Check network connectivity and firewall rules
5. Review database logs for errors

Slow Queries

• Optimize index configuration
• Reduce embedding dimensions
• Decrease TOP_K value
• Add more resources to database
• Enable caching

Out of Memory

• Use disk-based storage instead of in-memory
• Reduce batch size during indexing
• Scale database resources
• Archive old/unused collections

Best Practices

1. Choose Based on Scale:

   • Small datasets: ChromaDB
   • Medium: PGVector, Qdrant
   • Large: Milvus, Elasticsearch
   • Managed: Pinecone, Qdrant Cloud

2. Monitor Performance:

   • Track query latency
   • Monitor index size
   • Watch memory usage

3. Backup Strategy:

   • Regular database backups
   • Document source preservation
   • Embedding model versioning

4. Security:

   • Use authentication
   • Enable TLS/SSL
   • Network isolation
   • API key rotation

References

• ChromaDB: docs.trychroma.com
• Qdrant: qdrant.tech/documentation
• Milvus: milvus.io/docs
• PGVector: github.com/pgvector/pgvector