Google GenerativeAI SDK for .NET

To implement conversational AI, use the StartChat() method on a GenerativeModel instance. This creates a chat session that preserves context from previous messages, allowing for multi-turn reasoning. You can initialize the model using either GoogleAi or VertexAI providers.

Workflow:

1. Initialize the provider (GoogleAi or VertexAI).
2. Create a GenerativeModel using a model name (e.g., models/gemini-1.5-flash).
3. Call StartChat() to get a chat session object.
4. Use GenerateContentAsync(...) on the chat session to exchange messages.

// Using Google AI
var googleAI = new GoogleAi("YOUR_GOOGLE_API_KEY");
var generativeModel = googleAI.CreateGenerativeModel("models/gemini-1.5-flash");
var chatSession = generativeModel.StartChat();
var response = await chatSession.GenerateContentAsync("Hello!");

// Using Vertex AI
var vertexAI = new VertexAI(); 
var generativeModel = vertexAI.CreateGenerativeModel("models/gemini-1.5-flash");
var chatSession = generativeModel.StartChat();
var response = await chatSession.GenerateContentAsync("Hello!");

Gemini models support multimodal inputs including images, audio, and video. You can provide these files in three ways:

1. Local Files: Pass the absolute file path directly to GenerateContentAsync.
2. Remote Files: Pass the file URI and the corresponding MIME type to GenerateContentAsync.
3. GenerateContentRequest: For complex prompts, create a GenerateContentRequest object and use AddInlineFile(path) for local files or AddRemoteFile(uri, mimeType) for remote files.

// 1. Local File
var response = await geminiModel.GenerateContentAsync("Describe this image", @"C:\path\to\image.jpg");

// 2. Remote File
var response = await geminiModel.GenerateContentAsync("Summarize this PDF", "https://example.com/sample.pdf", "application/pdf");

// 3. Using GenerateContentRequest for multiple files
var request = new GenerateContentRequest();
request.AddText("Describe these files");
request.AddInlineFile(@"C:\files\example.png");
request.AddRemoteFile("https://example.com/path/to/sample.pdf", "application/pdf");
var response = await geminiModel.GenerateContentAsync(request);

The SDK provides several ways to work with structured JSON data from Gemini:

Automatic Deserialization

• GenerateObjectAsync<T>(request): Directly returns a deserialized object of type T.
• GenerateContentAsync<T>(request): Returns a response that can be converted using .ToObject<T>().
• UseJsonMode<T>(): An extension method for GenerateContentRequest that instructs the model to return a JSON response matching type T.

Manual JSON Parsing

If you need more control, you can manually configure the GenerationConfig with ResponseMimeType = "application/json" and a ResponseSchema. To extract data from a standard response, use response.ExtractJsonBlocks() to retrieve raw JSON blocks, then call .ToObject<T>() on each block.

// Option 1: Direct deserialization
var myObject = await model.GenerateObjectAsync<SampleJsonClass>(request);

// Option 2: Using JSON mode in request
var request = new GenerateContentRequest();
request.UseJsonMode<SampleJsonClass>();
request.AddText("Give me a response.");
var response = await model.GenerateContentAsync(request);
var myObject = response.ToObject<SampleJsonClass>();

// Option 3: Manual extraction from blocks
var response = await model.GenerateContentAsync(request);
var jsonBlocks = response.ExtractJsonBlocks();
var myObjects = jsonBlocks.Select(block => block.ToObject<SampleJsonClass>());

You can implement RAG by integrating the Vertex RAG Engine, which allows Gemini to access external corpora and vector databases (like Pinecone or Weaviate) for grounded responses.

To use this feature:

1. Initialize VertexAI with your platform configuration.
2. Use CreateRagManager() to get a manager for corpus operations.
3. Create a corpus using CreateCorpusAsync.
4. Import data into the corpus using ImportFilesAsync with a source (e.g., GcsSource).
5. Create a generative model using CreateGenerativeModel, passing the corpus.Name to the corpusIdForRag parameter to link the model to your knowledge base.
6. Call GenerateContentAsync to query the model using the retrieved context.

// Initialize VertexAI with your platform configuration.
var vertexAi = new VertexAI(GetTestVertexAIPlatform());

// Create an instance of the RAG manager for corpus operations.
var ragManager = vertexAi.CreateRagManager();

// Create a new corpus for your knowledge base.
var corpus = await ragManager.CreateCorpusAsync("My New Corpus", "My description");

// Import data into the corpus from a specified source.
var fileSource = new GcsSource() { /* Configure your GcsSource here */ };
await ragManager.ImportFilesAsync(corpus.Name, fileSource);

// Create a Gemini generative model configured to use the created corpus for RAG.
var model = vertexAi.CreateGenerativeModel(VertexAIModels.Gemini.Gemini2Flash, corpusIdForRag: corpus.Name);

// Generate content by querying the model.
var result = await model.GenerateContentAsync("query related to the corpus");