Chat Interfaces

Chat interfaces enable multi-turn conversations where the AI maintains context across messages. Genkit provides tools for managing conversation history and building stateful chat applications.

Basic Chat Flow

Create a simple chat flow that maintains message history:

import { genkit, z } from 'genkit';
import { googleAI } from '@genkit-ai/google-genai';
import type { MessageData } from '@genkit-ai/ai/model';

const ai = genkit({ plugins: [googleAI()] });

const ChatInputSchema = z.object({
  sessionId: z.string(),
  question: z.string(),
});

const ChatOutputSchema = z.object({
  sessionId: z.string(),
  history: z.array(MessageSchema),
});

// Simple in-memory storage for chat history
class ChatHistoryStore {
  private sessions: Map<string, MessageData[]> = new Map();
  
  write(sessionId: string, history: MessageData[]) {
    this.sessions.set(sessionId, history);
  }
  
  read(sessionId: string): MessageData[] {
    return this.sessions.get(sessionId) || [];
  }
}

const chatStore = new ChatHistoryStore();

const chatFlow = ai.defineFlow(
  {
    name: 'multiTurnChat',
    inputSchema: ChatInputSchema,
    outputSchema: ChatOutputSchema,
  },
  async (input) => {
    // Fetch chat history
    let history = await ai.run('fetchHistory', async () =>
      chatStore.read(input.sessionId)
    );

    // Generate response
    const llmResponse = await ai.generate({
      model: googleAI.model('gemini-2.5-flash'),
      messages: history,
      prompt: { text: input.question },
    });

    // Save updated history
    history = llmResponse.messages;
    chatStore.write(input.sessionId, history);
    
    return {
      sessionId: input.sessionId,
      history: history,
    };
  }
);

Message History Management

In-Memory Storage (Development)

Simple storage for development and testing:

export class ChatHistoryStore {
  private preamble: MessageData[];
  private sessions: Map<string, MessageData[]> = new Map();

  constructor(preamble: MessageData[] = []) {
    this.preamble = preamble;
  }

  write(sessionId: string, history: MessageData[]) {
    this.sessions.set(sessionId, history);
  }

  read(sessionId: string): MessageData[] {
    return this.sessions.get(sessionId) || this.preamble;
  }
}

Database Storage (Production)

For production, use a database like Firestore:

import { getFirestore } from 'firebase-admin/firestore';

class FirestoreChatStore {
  private db = getFirestore();
  
  async write(sessionId: string, history: MessageData[]) {
    await this.db.collection('chat_sessions').doc(sessionId).set({
      history,
      updatedAt: new Date(),
    });
  }
  
  async read(sessionId: string): Promise<MessageData[]> {
    const doc = await this.db.collection('chat_sessions').doc(sessionId).get();
    return doc.exists ? doc.data()?.history || [] : [];
  }
}

System Prompts and Preambles

Seed conversations with context and personality:

const menuData = loadMenuFromDatabase();

const preamble: MessageData[] = [
  {
    role: 'user',
    content: [{ text: "Hi. What's on the menu today?" }],
  },
  {
    role: 'model',
    content: [
      {
        text:
          'I am Walt, a helpful AI assistant here at the restaurant.\n' +
          'I can answer questions about the food on the menu.\n' +
          "Here is today's menu:\n" +
          menuData.map(item => `- ${item.title} ${item.price}\n${item.description}`).join('\n') +
          'Do you have any questions about the menu?',
      },
    ],
  },
];

const chatStore = new ChatHistoryStore(preamble);

Session Management

Genkit provides experimental session support for typed state across requests:

import "github.com/firebase/genkit/go/core/x/session"

type CartState struct {
    Items []string `json:"items"`
}

store := session.NewInMemoryStore[CartState]()

genkit.DefineFlow(g, "manageCart", 
    func(ctx context.Context, input string) (string, error) {
        // Load or create session
        sess, err := session.Load(ctx, store, "session-id")
        if err != nil {
            sess, _ = session.New(ctx,
                session.WithID[CartState]("session-id"),
                session.WithStore(store),
                session.WithInitialState(CartState{}),
            )
        }
        ctx = session.NewContext(ctx, sess)

        // Tools can access session state via session.FromContext[CartState](ctx)
        return genkit.GenerateText(ctx, g, 
            ai.WithPrompt(input), 
            ai.WithTools(cartTools...))
    })

See the session sample for a complete example.

Streaming Chat Responses

Stream chat responses in real-time:

const streamingChatFlow = ai.defineStreamingFlow(
  {
    name: 'streamingChat',
    inputSchema: ChatInputSchema,
    outputSchema: ChatOutputSchema,
  },
  async (input, { streamingCallback }) => {
    const history = chatStore.read(input.sessionId);

    const { messages, stream } = await ai.generate({
      model: googleAI.model('gemini-2.5-flash'),
      messages: history,
      prompt: { text: input.question },
      streamingCallback,
    });

    chatStore.write(input.sessionId, messages);
    
    return {
      sessionId: input.sessionId,
      history: messages,
    };
  }
);

Context Window Management

Manage conversation length to stay within model limits:

function truncateHistory(
  history: MessageData[], 
  maxMessages: number = 20
): MessageData[] {
  // Keep system message and last N messages
  const systemMessages = history.filter(m => m.role === 'system');
  const recentMessages = history
    .filter(m => m.role !== 'system')
    .slice(-maxMessages);
  
  return [...systemMessages, ...recentMessages];
}

const chatFlow = ai.defineFlow(
  { name: 'chat', inputSchema: ChatInputSchema },
  async (input) => {
    let history = chatStore.read(input.sessionId);
    
    // Truncate to prevent exceeding context window
    history = truncateHistory(history, 20);
    
    const llmResponse = await ai.generate({
      model: googleAI.model('gemini-2.5-flash'),
      messages: history,
      prompt: { text: input.question },
    });
    
    chatStore.write(input.sessionId, llmResponse.messages);
    return llmResponse.messages;
  }
);

Conversation Summarization

Summarize old messages to save context:

async function summarizeHistory(
  history: MessageData[]
): Promise<MessageData[]> {
  if (history.length < 10) return history;
  
  // Summarize older messages
  const oldMessages = history.slice(0, -5);
  const recentMessages = history.slice(-5);
  
  const summary = await ai.generate({
    model: googleAI.model('gemini-2.5-flash'),
    prompt: `Summarize this conversation concisely:\n${JSON.stringify(oldMessages)}`,
  });
  
  return [
    { role: 'system', content: [{ text: `Previous conversation summary: ${summary.text}` }] },
    ...recentMessages,
  ];
}

Multi-User Chat

Handle multiple concurrent users:

class MultiUserChatStore {
  private sessions: Map<string, Map<string, MessageData[]>> = new Map();
  
  write(userId: string, sessionId: string, history: MessageData[]) {
    if (!this.sessions.has(userId)) {
      this.sessions.set(userId, new Map());
    }
    this.sessions.get(userId)!.set(sessionId, history);
  }
  
  read(userId: string, sessionId: string): MessageData[] {
    return this.sessions.get(userId)?.get(sessionId) || [];
  }
  
  listSessions(userId: string): string[] {
    return Array.from(this.sessions.get(userId)?.keys() || []);
  }
}

Best Practices

Generate Unique Session IDs

Use UUIDs or similar for session identification:

TypeScript

import { v4 as uuidv4 } from 'uuid';

const sessionId = uuidv4();

Set Clear System Instructions

Define the assistant’s role and boundaries:

TypeScript

const systemMessage = {
  role: 'system',
  content: [{
    text: 'You are a helpful restaurant assistant. Only answer questions about the menu. For other topics, politely redirect to menu-related questions.'
  }]
};

Implement Session Timeouts

Clear old sessions to save storage:

TypeScript

class ChatStoreWithTTL {
  private sessions: Map<string, { history: MessageData[], lastAccess: Date }> = new Map();
  private ttlMinutes = 60;
  
  write(sessionId: string, history: MessageData[]) {
    this.sessions.set(sessionId, {
      history,
      lastAccess: new Date(),
    });
    this.cleanupOldSessions();
  }
  
  private cleanupOldSessions() {
    const now = new Date();
    for (const [id, session] of this.sessions.entries()) {
      const age = (now.getTime() - session.lastAccess.getTime()) / 1000 / 60;
      if (age > this.ttlMinutes) {
        this.sessions.delete(id);
      }
    }
  }
}

Handle Message Validation

Validate and sanitize user inputs:

TypeScript

function validateMessage(message: string): string {
  // Remove excessive whitespace
  message = message.trim().replace(/\s+/g, ' ');
  
  // Check length
  if (message.length === 0) {
    throw new Error('Message cannot be empty');
  }
  if (message.length > 4000) {
    throw new Error('Message too long');
  }
  
  return message;
}

Complete Chat Application Example

import { genkit, z } from 'genkit';
import { googleAI } from '@genkit-ai/google-genai';
import type { MessageData } from '@genkit-ai/ai/model';

const ai = genkit({ plugins: [googleAI()] });

// Load menu data
const menuData = require('./menu.json');

// System preamble
const preamble: MessageData[] = [
  {
    role: 'user',
    content: [{ text: "Hi. What's on the menu?" }],
  },
  {
    role: 'model',
    content: [
      {
        text:
          'I am Walt, a helpful AI assistant at the restaurant.\n' +
          'I can answer questions about our menu.\n' +
          "Here's what we have today:\n" +
          menuData.map(r => `- ${r.title} ${r.price}\n${r.description}`).join('\n'),
      },
    ],
  },
];

const chatStore = new ChatHistoryStore(preamble);

const chatFlow = ai.defineFlow(
  {
    name: 'restaurantChat',
    inputSchema: z.object({
      sessionId: z.string(),
      question: z.string(),
    }),
    outputSchema: z.object({
      sessionId: z.string(),
      history: z.array(MessageSchema),
    }),
  },
  async (input) => {
    // Fetch history
    let history = await ai.run('fetchHistory', async () =>
      chatStore.read(input.sessionId)
    );

    // Generate response
    const llmResponse = await ai.generate({
      model: googleAI.model('gemini-2.5-flash'),
      messages: history,
      prompt: { text: input.question },
    });

    // Save history
    history = llmResponse.messages;
    chatStore.write(input.sessionId, history);
    
    return {
      sessionId: input.sessionId,
      history: history,
    };
  }
);

export { chatFlow };

Next Steps

Learn about Tool Calling to add capabilities
Explore Streaming for real-time responses
Check out Evaluation to test chat quality

Overview

Getting Started

Core Concepts

Guides

Model Providers

Deployment

Developer Tools

Chat Interfaces

Chat Interfaces

Basic Chat Flow

Message History Management

In-Memory Storage (Development)

Database Storage (Production)

System Prompts and Preambles

Session Management

Streaming Chat Responses

Context Window Management

Conversation Summarization

Multi-User Chat

Best Practices

Generate Unique Session IDs

Set Clear System Instructions

Implement Session Timeouts

Handle Message Validation

Complete Chat Application Example

Next Steps

Build docs developers (and LLMs) love

Overview

Getting Started

Core Concepts

Guides

Model Providers

Deployment

Developer Tools

Documentation Index

​Chat Interfaces

​Basic Chat Flow

​Message History Management

​In-Memory Storage (Development)

​Database Storage (Production)

​System Prompts and Preambles

​Session Management

​Streaming Chat Responses

​Context Window Management

​Conversation Summarization

​Multi-User Chat

​Best Practices

​Generate Unique Session IDs

​Set Clear System Instructions

​Implement Session Timeouts

​Handle Message Validation

​Complete Chat Application Example

​Next Steps

Build docs developers (and LLMs) love

Chat Interfaces

Basic Chat Flow

Message History Management

In-Memory Storage (Development)

Database Storage (Production)

System Prompts and Preambles

Session Management

Streaming Chat Responses

Context Window Management

Conversation Summarization

Multi-User Chat

Best Practices

Generate Unique Session IDs

Set Clear System Instructions

Implement Session Timeouts

Handle Message Validation

Complete Chat Application Example

Next Steps