TiDB.Vector.AzureOpenAI 1.3.0

TiDB.Vector.NET

Ergonomic C# SDK for TiDB Vector Search: upsert, search, and RAG with a fluent builder, OpenAI embeddings/chat integration, and DX-first defaults.

Repository: manasseh-zw/TiDB.Vector.NET

Installation

# Core package with OpenAI built-in (recommended)
dotnet add package TiDB.Vector

# Azure OpenAI integration (optional, extends core)
dotnet add package TiDB.Vector.AzureOpenAI

Features

• Fluent builder to wire connection, embeddings, and chat
• Default schema with dedicated tags JSON column for efficient filtering
• Safe, parameterized SQL via MySqlConnector
• OpenAI providers built into the core package
• Simple RAG helper (AskAsync) that cites sources
• Advanced filtering: Collection filtering and key-value tag filtering
• Source tracking: Track document sources (URLs, file paths, etc.)
• Samples with .env support via dotenv.net

NuGet Packages

• TiDB.Vector - Core package with vector store functionality

• TiDB.Vector.AzureOpenAI - Azure OpenAI integration (embeddings + chat)

Projects

• TiDB.Vector (Core): store API, schema, SQL

• TiDB.Vector.AzureOpenAI: Azure OpenAI providers (embeddings/chat)

• TiDB.Vector.Samples: runnable examples (upsert, search, ask)

Requirements

• .NET 8.0+
• TiDB v8.4+ (v8.5+ recommended); TiDB Cloud supported
• For HNSW index: TiFlash replica required on the target table

Installation

# Core package with OpenAI built-in (recommended)
dotnet add package TiDB.Vector

# Azure OpenAI integration (optional, extends core)
dotnet add package TiDB.Vector.AzureOpenAI

What You Get With Each Package

Note: The core package now provides everything you need! Azure OpenAI is an optional extension for Azure-specific features.

Quickstart (local dev)

1. Clone the repo:

git clone https://github.com/manasseh-zw/TiDB.Vector.NET
cd TiDB.Vector.NET

2. Create a .env file in TiDB.Vector.Samples/:

# Required for all samples
TIDB_CONN_STRING=Server=<host>;Port=4000;User ID=<user>;Password=<pwd>;Database=<db>;SslMode=VerifyFull;

# For OpenAI samples
OPENAI_API_KEY=sk-...

# For Azure OpenAI samples
AZURE_AI_APIKEY=your-azure-key
AZURE_AI_ENDPOINT=https://your-resource.openai.azure.com/

3. Run samples:

dotnet run --project TiDB.Vector.Samples

The samples include:

• FilteringSample: Demonstrates collection filtering and tag-based filtering
• SearchSample: Basic vector search functionality
• AskSample: RAG (Retrieval-Augmented Generation) with source citations
• UpsertSample: Document ingestion with chunking
• AzureOpenAISample: Azure OpenAI integration example

Each sample will:

• Ensure the default schema and (optionally) HNSW index
• Upsert sample documents with tags and sources
• Perform filtered vector searches
• Call AskAsync to answer questions using retrieved context

Core API (glance)

New Clean API (Recommended)

using TiDB.Vector.Core;
using TiDB.Vector.OpenAI.Builder; // Built into TiDB.Vector core package

var store = new TiDBVectorStoreBuilder(
        Environment.GetEnvironmentVariable("TIDB_CONN_STRING")!)
    .WithDefaultCollection("docs")
    .WithDistanceFunction(DistanceFunction.Cosine)
    .AddOpenAI(Environment.GetEnvironmentVariable("OPENAI_API_KEY")!)
    .AddOpenAITextEmbedding("text-embedding-3-small", 1536)
    .AddOpenAIChatCompletion("gpt-4o-mini")
    .EnsureSchema(createVectorIndex: true)
    .Build();

Custom Endpoints (OpenAI-Compatible Services)

var store = new TiDBVectorStoreBuilder(connectionString)
    .AddOpenAI("your-api-key", "https://your-local-model.com/v1")
    .AddOpenAITextEmbedding("your-embedding-model", 1536)
    .AddOpenAIChatCompletion("your-chat-model")
    .Build();

Azure OpenAI Integration

var store = new TiDBVectorStoreBuilder(connectionString)
    .AddAzureOpenAI("your-azure-key", "https://your-resource.openai.azure.com/")
    .AddAzureOpenAITextEmbedding("your-embedding-deployment", 1536)
    .AddAzureOpenAIChatCompletion("your-chat-deployment")
    .Build();

await store.EnsureSchemaAsync();

await store.UpsertAsync(new UpsertItem
{
Id = "sample-1",
Collection = "docs",
Content = "Fish live in water and are known for their swimming abilities."
});

var results = await store.SearchAsync("a swimming animal", topK: 3);
var answer = await store.AskAsync("Name an animal that swims.");

Advanced Filtering

The SDK supports powerful filtering capabilities for multi-tenant and fine-grained search:

Collection Filtering

Filter by a specific collection (one at a time for simplicity):

var results = await store.SearchAsync(
    "query",
    searchOptions: new SearchOptions { Collection = "engineering-docs" }
);

Tag Filtering (New DX)

Type-safe tag filters using Tag and TagFilter:

using TiDB.Vector.Models;

var results = await store.SearchAsync(
    "query",
    searchOptions: new SearchOptions
    {
        TagFilter = new TagFilter(new[]
        {
            new Tag("OrganizationId", "org-123"),
            new Tag("Department", "Engineering")
        }, TagFilterMode.And)
    }
);

Combined Filtering

var results = await store.SearchAsync(
    "query",
    searchOptions: new SearchOptions
    {
        Collection = "engineering-docs",
        TagFilter = new TagFilter(new[]
        {
            new Tag("OrganizationId", "org-123"),
            new Tag("Department", "Engineering")
        }, TagFilterMode.And)
    }
);

Document Structure with Tags

When upserting documents, use the dedicated tags column with Tag:

await store.UpsertAsync(new UpsertItem
{
    Id = "doc-1",
    Collection = "engineering-docs",
    Content = "How to implement microservices...",
    Source = "https://docs.company.com/guide.pdf",
    Metadata = JsonDocument.Parse("""
    {
        "Category": "Architecture",
        "Tags": ["microservices", "docker"]
    }
    """),
    Tags = new[] { new Tag("OrganizationId", "org-123"), new Tag("Department", "Engineering") }
});

RAG with Filtering

Use filtering in RAG queries:

var answer = await store.AskAsync(
    "What are the best practices for software development?",
    topK: 3,
    searchOptions: new SearchOptions
    {
        Collection = "engineering-docs",
        TagFilter = new Tag("Department", "Engineering")
    }
);

Azure OpenAI Example

using TiDB.Vector.Core;
using TiDB.Vector.AzureOpenAI.Builder;

var store = new TiDBVectorStoreBuilder(
        Environment.GetEnvironmentVariable("TIDB_CONN_STRING")!)
    .WithDefaultCollection("docs")
    .WithDistanceFunction(DistanceFunction.Cosine)
    .AddAzureOpenAI(
        Environment.GetEnvironmentVariable("AZURE_AI_APIKEY")!,
        Environment.GetEnvironmentVariable("AZURE_AI_ENDPOINT")!)
    .AddAzureOpenAITextEmbedding("your-embedding-deployment", 1536)
    .AddAzureOpenAIChatCompletion("your-chat-deployment")
    .EnsureSchema(createVectorIndex: true)
    .Build();

Default Schema and Index

• Table: tidb_vectors
  • collection VARCHAR(128) + id VARCHAR(64) as composite primary key
  • embedding VECTOR (variable dimensions supported)
  • content TEXT, metadata JSON, source VARCHAR(512), tags JSON, timestamps
• Index (optional, HNSW):
  • Cosine: CREATE VECTOR INDEX idx_tidb_vectors_embedding_cosine ON tidb_vectors ((VEC_COSINE_DISTANCE(embedding))) USING HNSW;
  • L2: CREATE VECTOR INDEX idx_tidb_vectors_embedding_l2 ON tidb_vectors ((VEC_L2_DISTANCE(embedding))) USING HNSW;

Schema Features:

• Dedicated tags column: Efficient JSON-based filtering without parsing metadata
• Source tracking: Track document origins (URLs, file paths, etc.)
• Variable vector dimensions: Support for different embedding models
• TiFlash replica required for building/using the HNSW index
• Optimized filtering: KNN first, then apply filters to maintain index usage

Filtering Performance:

• Dedicated tags column: Faster than parsing metadata JSON for filtering
• TiDB JSON optimization: Leverages TiDB's binary JSON serialization for quick access
• Index-friendly: Maintains vector index usage during filtered searches
• Multi-tenant ready: Efficient organization/department filtering for SaaS applications

OpenAI integration

Built into the core package using the official OpenAI .NET SDK 2.x:

• Embeddings: EmbeddingClient (text-embedding-3-small 1536 dims, text-embedding-3-large 3072 dims)
• Chat: ChatClient (e.g., gpt-4o-mini)

Ensure the table dimension matches the chosen embedding model’s dimension.

Connection string (TiDB Cloud)

Use standard ADO.NET format (not URL form), e.g.:

Server=<host>;Port=4000;User ID=<user>;Password=<pwd>;Database=<db>;SslMode=VerifyFull;

If you must provide a custom CA bundle, append:

SslCa=C:\\path\\to\\isrgrootx1.pem;

Current Architecture

The project now has OpenAI integration built into the core package:

• TiDB.Vector - Core vector store functionality + OpenAI integration built-in
• TiDB.Vector.AzureOpenAI - Azure OpenAI integration (extends core capabilities)

Note: Users now get full AI functionality with just the core package! Azure OpenAI is available as an optional extension.

Future Development Plans

We're planning to restructure the architecture for better user experience:

Phase 1: OpenAI Built-In ✅ COMPLETED

• TiDB.Vector - Core + OpenAI integration built-in by default
• Users get AI capabilities out of the box with one package
• OpenAI-compatible endpoint support for local models and other providers
• Provider abstraction layer for easy integration with OpenAI-compatible services

Phase 2: Provider Extensions

• TiDB.Vector.AzureOpenAI - Azure-specific optimizations
• TiDB.Vector.GoogleGemini - Google Gemini integration
• TiDB.Vector.Anthropic - Anthropic Claude integration
• TiDB.Vector.Custom - Template for custom provider implementations

Benefits of New Architecture

• ✅ One package = full functionality
• ✅ AI works immediately without additional packages
• ✅ Easy provider switching via configuration
• ✅ Future-proof with clean extension points
• ✅ Better developer experience

Roadmap (high level)

• Iteration 1: Skeleton and builder ✅
• Iteration 2: Schema management + upsert/search ✅
• Iteration 3: Vector index + TiFlash helpers ✅
• Iteration 4: Chunking (out-of-the-box text splitter) ✅
• Iteration 5: OpenAI embeddings/chat + RAG AskAsync ✅
• Iteration 6: Azure OpenAI integration + additional providers ✅
• Iteration 7: OpenAI built into core + provider abstraction layer ✅ COMPLETED
• Iteration 8: Advanced filtering (collection + tag filtering) + source tracking ✅ COMPLETED
• Iteration 9: Additional provider extensions (Gemini, Claude, etc.)
• Iteration 10+: Advanced features (metrics, streaming, hybrid search, performance optimizations)

Contributing

We welcome issues and PRs! To contribute:

• Fork the repo and create a feature branch
• Follow the existing code style (explicit naming, clear control flow)
• Keep public APIs strongly-typed and documented
• Add/update samples if you change user-facing behavior
• Run dotnet build before submitting your PR

Open an issue to discuss larger proposals or provider integrations. Repo: manasseh-zw/TiDB.Vector.NET

Acknowledgements

• TiDB Vector Search (data types, functions, HNSW index)
• Official OpenAI .NET SDK for embeddings/chat