Thoth Vector Database Manager v0.6.2

A high-performance, Haystack v2-based vector database manager with external embedding providers and centralized embedding management for 4 production-ready backends.

🤖 MCP Server Support

This project is configured with MCP (Model Context Protocol) servers for enhanced AI-assisted development:

• Context7: Enhanced context management
• Serena: IDE assistance and development support

See docs/MCP_SETUP.md for details.

🚀 Features

🌐 NEW in v0.6.0: External Embedding Providers

• OpenAI, Cohere, Mistral: Support for major external embedding APIs
• Cost-Effective: Pay-per-use model with intelligent caching
• High-Quality Embeddings: State-of-the-art embedding models
• Unified Management: Centralized ExternalEmbeddingManager

🏗️ Core Features

• Multi-backend support: Qdrant, Chroma, PostgreSQL pgvector, Milvus
• Haystack v2 integration: Uses haystack-ai v2.12.0+ as an abstraction layer
• Centralized embeddings: No more client-side embedding management
• Memory optimization: Intelligent caching and lazy loading
• API compatibility: Backward compatible with existing APIs
• Type safety: Full type hints and Pydantic validation
• Production-ready: Comprehensive testing and robust error handling

📦 Installation

🚀 Recommended: uv Package Manager

This project uses uv for fast, reliable Python package management. Install uv first:

# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

✅ No Dependency Conflicts

Version 0.4.0 resolves all dependency conflicts! All 4 supported databases can now be installed together:

All Databases (Recommended)

# Install all supported backends (Qdrant, Chroma, PgVector, Milvus)
uv add thoth-vdbmanager[all]

Individual Backends

# Individual backend installation
uv add thoth-vdbmanager[qdrant]    # Qdrant support
uv add thoth-vdbmanager[chroma]    # Chroma support
uv add thoth-vdbmanager[pgvector]  # PostgreSQL pgvector support
uv add thoth-vdbmanager[milvus]    # Milvus support

Development Installation

# For development with all backends and testing tools
uv add thoth-vdbmanager[all,test,dev]

🔄 pip Installation (Also Supported)

If you prefer pip, all commands work by replacing uv add with pip install:

# Example with pip
pip install thoth-vdbmanager[all]

🔄 Breaking Changes in v0.4.0

• Removed: Weaviate and Pinecone support (no longer maintained)
• Updated: Now requires haystack-ai v2.12.0+ (not compatible with legacy haystack)
• Improved: All remaining databases work together without conflicts

🏗️ Architecture

The library is built on a clean architecture with:

• Core: Base interfaces and document types
• Adapters: Backend-specific implementations using Haystack
• Factory: Unified creation interface
• Compatibility: Legacy API support

🌐 External Embedding Providers (NEW in v0.6.0)

Setup External Embeddings

Configure your external embedding provider using environment variables:

# OpenAI (recommended)
export EMBEDDING_PROVIDER=openai
export EMBEDDING_API_KEY=sk-your-openai-key
export EMBEDDING_MODEL=text-embedding-3-small

# Cohere
export EMBEDDING_PROVIDER=cohere  
export EMBEDDING_API_KEY=your-cohere-key
export EMBEDDING_MODEL=embed-multilingual-v3.0

# Mistral
export EMBEDDING_PROVIDER=mistral
export EMBEDDING_API_KEY=your-mistral-key
export EMBEDDING_MODEL=mistral-embed

Using External Embeddings

import os
from thoth_vdbmanager import ExternalVectorStoreFactory, ColumnNameDocument

# Create store with external embeddings
store = ExternalVectorStoreFactory.create_from_env(
    backend="qdrant",
    collection="my_collection",
    host="localhost",
    port=6333
)

# Add document - embeddings generated via API
doc = ColumnNameDocument(
    table_name="users",
    column_name="email",
    column_description="User email address",
    value_description="Valid email format"
)
store.add_column_description(doc)

# Search - query embeddings generated via API
results = store.search_similar(
    query="user email address",
    doc_type="column_name", 
    top_k=5
)

Available External Providers

Cost Optimization with Caching

# Enable intelligent caching to reduce API calls
embedding_config = {
    'provider': 'openai',
    'api_key': 'sk-your-key',
    'model': 'text-embedding-3-small',
    'enable_cache': True,    # Enable caching
    'cache_size': 10000      # Cache up to 10k embeddings
}

store = ExternalVectorStoreFactory.create(
    backend="qdrant",
    embedding_config=embedding_config,
    collection="cached_collection",
    host="localhost",
    port=6333
)

🚀 Quick Start

External Embedding API (Recommended)

import os
from thoth_vdbmanager import ExternalVectorStoreFactory, ColumnNameDocument, SqlDocument, EvidenceDocument

# Set up external embedding provider
os.environ['EMBEDDING_PROVIDER'] = 'openai'
os.environ['EMBEDDING_API_KEY'] = 'sk-your-openai-key'
os.environ['EMBEDDING_MODEL'] = 'text-embedding-3-small'

# Create a vector store with external embeddings
store = ExternalVectorStoreFactory.create_from_env(
    backend="qdrant",
    collection="my_collection",
    host="localhost",
    port=6333
)

# Add documents
column_doc = ColumnNameDocument(
    table_name="users",
    column_name="email",
    original_column_name="user_email",
    column_description="User email address",
    value_description="Valid email format"
)

doc_id = store.add_column_description(column_doc)

# Search documents using external API embeddings
results = store.search_similar(
    query="user email",
    doc_type="column_name",
    top_k=5
)

Available Classes

from thoth_vdbmanager import (
    VectorStoreFactory,      # Main factory for creating stores
    ColumnNameDocument,      # Column metadata documents
    SqlDocument,            # SQL example documents
    EvidenceDocument,       # Evidence/hint documents
    ThothType,              # Document type enumeration
    VectorStoreInterface    # Base interface for all stores
)

🔧 Configuration

Qdrant

store = VectorStoreFactory.create(
    backend="qdrant",
    collection="my_collection",
    host="localhost",
    port=6333,
    api_key="your-api-key",  # Optional
    embedding_dim=384,  # Optional
    hnsw_config={"m": 16, "ef_construct": 100}
)

Chroma (Multiple Modes)

Memory Mode (Recommended for Testing):

store = VectorStoreFactory.create(
    backend="chroma",
    collection="my_collection",
    mode="memory"  # Fast, isolated, no persistence
)

Filesystem Mode:

store = VectorStoreFactory.create(
    backend="chroma",
    collection="my_collection",
    mode="filesystem",
    persist_path="./chroma_db"
)

Server Mode (Production):

store = VectorStoreFactory.create(
    backend="chroma",
    collection="my_collection",
    mode="server",
    host="localhost",
    port=8000
)

📖 See Chroma Configuration Guide for detailed setup instructions

PostgreSQL pgvector

store = VectorStoreFactory.create(
    backend="pgvector",
    collection="my_table",
    connection_string="postgresql://user:pass@localhost:5432/dbname"
)

Milvus (Multiple Modes)

Lite Mode (Recommended for Testing):

store = VectorStoreFactory.create(
    backend="milvus",
    collection="my_collection",
    mode="lite",
    connection_uri="./milvus.db"  # File-based storage
)

Server Mode (Production):

store = VectorStoreFactory.create(
    backend="milvus",
    collection="my_collection",
    mode="server",
    host="localhost",
    port=19530
)

📖 See Milvus Configuration Guide for detailed setup instructions

📊 Performance Optimizations

Memory Usage

• Lazy initialization: Embedders and connections are initialized on first use
• Singleton pattern: Same configuration reuses existing instances
• Batch processing: Efficient bulk operations

Performance Tuning

# Optimize for specific use cases
store = VectorStoreFactory.create(
    backend="qdrant",
    collection="optimized",
    embedding_model="sentence-transformers/all-MiniLM-L6-v2",  # 384-dim, fast
    hnsw_config={"m": 32, "ef_construct": 200}  # Better search quality
)

🧪 Testing

# Run all tests
pytest

# Run specific backend tests
pytest tests/test_qdrant.py -v

# Run with coverage
pytest --cov=vdbmanager tests/

📈 Migration Guide

From v0.3.x to v0.4.0

Breaking Changes

• Removed databases: Weaviate and Pinecone are no longer supported
• Haystack version: Now requires haystack-ai v2.12.0+ (not compatible with legacy haystack)
• Dependencies: All remaining databases can now be installed together without conflicts

Migration Steps

1. Update installation:

# Old installation (v0.3.x)
pip install thoth-vdbmanager[all-safe]  # Avoided conflicts

# New installation (v0.4.0)
pip install thoth-vdbmanager[all]  # No conflicts!

2. Update code (if using removed databases):

# If you were using Weaviate - migrate to Qdrant or Chroma
# Old code (v0.3.x)
store = VectorStoreFactory.create(
    backend="weaviate",  # No longer supported
    collection="MyCollection",
    url="http://localhost:8080"
)

# New code (v0.4.0) - migrate to similar database
store = VectorStoreFactory.create(
    backend="qdrant",  # Recommended alternative
    collection="my_collection",
    host="localhost",
    port=6333
)

3. Existing supported databases work unchanged:

# This code works exactly the same in v0.4.0
store = VectorStoreFactory.create(
    backend="qdrant",  # ✅ Still supported
    collection="my_docs",
    host="localhost",
    port=6333
)

🔍 API Reference

Core Classes

VectorStoreFactory

# Create store
store = VectorStoreFactory.create(backend, collection, **kwargs)

# From config
config = {"backend": "qdrant", "params": {...}}
store = VectorStoreFactory.from_config(config)

# List backends
backends = VectorStoreFactory.list_backends()

Document Types

• ColumnNameDocument: Column metadata
• SqlDocument: SQL examples
• EvidenceDocument: General evidence/hints

Methods

• add_column_description(doc): Add column metadata
• add_sql(doc): Add SQL example
• add_evidence(doc): Add evidence/hint
• search_similar(query, doc_type, top_k=5, score_threshold=0.7): Semantic search
• get_document(doc_id): Retrieve by ID
• bulk_add_documents(docs): Batch insert
• get_collection_info(): Get stats

🐛 Troubleshooting

Common Issues

Connection Errors

# Check service availability
import requests
requests.get("http://localhost:6333")  # Qdrant

Memory Issues

# Use smaller embedding model
store = VectorStoreFactory.create(
    backend="qdrant",
    collection="my_collection",
    embedding_model="sentence-transformers/all-MiniLM-L6-v2"  # 384-dim
)

Performance Issues

# Tune HNSW parameters
store = VectorStoreFactory.create(
    backend="qdrant",
    collection="my_collection",
    hnsw_config={"m": 16, "ef_construct": 100}
)

🤝 Contributing

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Ensure all tests pass
5. Submit a pull request

📄 License

MIT License - see LICENSE file for details.

📁 Directory Structure

thoth_vdbmanager/
├── vdbmanager/
│   ├── core/                    # Base interfaces and document types
│   │   ├── base.py             # Core document classes and interfaces
│   │   └── __init__.py
│   ├── adapters/               # Backend-specific implementations
│   │   ├── haystack_adapter.py # Base Haystack adapter
│   │   ├── qdrant_adapter.py   # Qdrant implementation
│   │   ├── chroma_adapter.py   # Chroma implementation
│   │   ├── pgvector_adapter.py # PostgreSQL pgvector
│   │   └── milvus_adapter.py   # Milvus implementation
│   ├── factory.py              # Unified creation interface
│   └── __init__.py            # Public API exports
├── test_e2e_vectordb/          # End-to-end tests
├── pyproject.toml              # Project configuration
└── README.md                   # This file

🚀 Quick API Reference

Main API

from thoth_vdbmanager import VectorStoreFactory, ColumnNameDocument

# Create any backend
store = VectorStoreFactory.create(
    backend="qdrant",
    collection="my_docs",
    host="localhost",
    port=6333
)

# Use the methods
doc_id = store.add_column_description(column_doc)
results = store.search_similar("user email", "column_name")

🎉 Ready to use with Haystack v2 and 4 production-ready vector databases!