Zur Interactions API migrieren

In dieser Anleitung erfahren Sie, wie Sie von der generateContent API zur Interactions API migrieren.

Die Interactions API ist die Standardschnittstelle für die Entwicklung mit Gemini. Sie ist für Agenten-Workflows, serverseitige Statusverwaltung und komplexe multimodale Unterhaltungen über mehrere Themen optimiert, unterstützt aber weiterhin einfache zustandslose Anfragen über ein einzelnes Thema. generateContent wird weiterhin vollständig unterstützt. Wir empfehlen jedoch die Interactions API für alle neuen Entwicklungen.

Warum migrieren?

Die Interactions API bietet eine strukturiertere und leistungsstärkere Möglichkeit, mit Gemini zu entwickeln:

  • Serverseitige Verlaufsverwaltung: Vereinfachte Abläufe über mehrere Themen mit previous_interaction_id. Der Server aktiviert den Status standardmäßig (store=true). Sie können jedoch zustandsloses Verhalten aktivieren, indem Sie store=false festlegen.
  • Beobachtbare Ausführungsschritte: Mit typisierten Schritten lassen sich komplexe Abläufe einfach debuggen und die Benutzeroberfläche für Zwischenereignisse (z. B. Gedanken oder Such-Widgets) rendern.
  • Für Agenten-Workflows entwickelt: Native Unterstützung für die mehrstufige Toolnutzung, Orchestrierung und komplexe Schlussfolgerungsabläufe durch typisierte Ausführungsschritte.
  • Lang andauernde und Hintergrundaufgaben: Unterstützt das Auslagern zeitaufwendiger Vorgänge wie Deep Think und Deep Research an Hintergrundprozesse mit background=true.
  • Zugriff auf neue Modelle und Funktionen: Künftig werden neue Modelle, die über die Hauptmodellfamilie hinausgehen, sowie neue agentische Funktionen und ‑tools ausschließlich in der Interactions API eingeführt. generateContent wird weiterhin vollständig für bestehende Anwendungsfälle unterstützt.

Einfache Eingabe/Ausgabe

In diesem Abschnitt wird gezeigt, wie Sie eine einfache Anfrage zur Textgenerierung migrieren.

Vorher (generateContent)

Die generateContent API ist zustandslos und gibt die Antwort direkt zurück. Die Antwortstruktur umschließt die Ausgabe in einer Liste von candidates, die jeweils content mit einer Liste von zu parsierenden parts enthalten.

Python

from google import genai

client = genai.Client()

response = client.models.generate_content(
    model="gemini-2.5-flash", contents="Tell me a joke."
)
print(response.text)

JavaScript

import { GoogleGenAI } from '@google/genai';

const ai = new GoogleGenAI({});

const response = await ai.models.generateContent({
  model: "gemini-2.5-flash",
  contents: "Tell me a joke.",
});
console.log(response.text);

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [{
        "parts": [{
            "text": "Tell me a joke."
        }]
    }]
}'

# Response
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "Why did the chicken cross the road? To get to the other side!"
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP",
      "index": 0
    }
  ],
  "usageMetadata": {
    "promptTokenCount": 4,
    "candidatesTokenCount": 12,
    "totalTokenCount": 16
  }
}

Nachher (Interactions API)

Die Interactions API gibt eine gespeicherte Interaktionsressource mit einer steps-Zeitachse zurück. Anstatt Kandidaten und Teile zu durchlaufen, prüfen Sie das steps-Array, um den gewünschten Ausgabetyp zu finden.

Python

from google import genai

client = genai.Client()

# The input can be a simple string shorthand
interaction = client.interactions.create(
    model="gemini-3-flash-preview", input="Tell me a joke."
)

# Inspect the steps manually
for step in interaction.steps:
    if step.type == "model_output":
        print(step.content[0].text)

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

let interaction = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: 'Tell me a joke.'
});

// Manual inspection
const modelStep = interaction.steps.find(s => s.type === 'model_output');
console.log(modelStep.content[0].text);

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "input": "Tell me a joke."
}'

# Response
{
  "id": "int_123",
  "status": "completed",
  "steps": [
    {
      "type": "user_input",
      "status": "done",
      "content": [
        {
          "type": "text",
          "text": "Tell me a joke."
        }
      ]
    },
    {
      "type": "model_output",
      "status": "done",
      "content": [
        {
          "type": "text",
          "text": "Why did the chicken cross the road?"
        }
      ]
    }
  ]
}

Unterhaltungen über mehrere Themen

Die Interactions API speichert Interaktionen standardmäßig und ermöglicht so die serverseitige Statusverwaltung für Unterhaltungen über mehrere Themen.

Vorher (generateContent)

In generateContent müssen Sie den Unterhaltungsverlauf manuell mit dem contents-Array oder einem clientseitigen Chat-Helfer verwalten.

Python

Chat-Helfer verwenden (empfohlen)

from google import genai

client = genai.Client()

chat = client.chats.create(model="gemini-2.5-flash")
response1 = chat.send_message("Hi, my name is Phil.")
print(response1.text)

response2 = chat.send_message("What is my name?")
print(response2.text)

Verlauf manuell verwalten

from google import genai
from google.genai import types

client = genai.Client()

# The second turn requires sending the entire history
response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents=[
        types.Content(
            role="user", parts=[types.Part.from_text("Hi, my name is Phil.")]
        ),
        types.Content(
            role="model",
            parts=[types.Part.from_text("Hi Phil, how can I help you?")],
        ),
        types.Content(
            role="user", parts=[types.Part.from_text("What is my name?")]
        ),
    ],
)
print(response.text)

JavaScript

Chat-Helfer verwenden (empfohlen)

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const chat = client.chats.create({ model: 'gemini-2.5-flash' });
let response = await chat.sendMessage({ message: 'Hi, my name is Phil.' });
console.log(response.text);

response = await chat.sendMessage({ message: 'What is my name?' });
console.log(response.text);

Verlauf manuell verwalten

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

// The second turn requires sending the entire history
const response = await client.models.generateContent({
    model: 'gemini-2.5-flash',
    contents: [
        { role: 'user', parts: [{ text: 'Hi, my name is Phil.' }] },
        { role: 'model', parts: [{ text: 'Hi Phil, how can I help you?' }] },
        { role: 'user', parts: [{ text: 'What is my name?' }] }
    ]
});
console.log(response.text);

REST

# Request (the second turn requires sending the entire history)
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [
        {"role": "user", "parts": [{"text": "Hi, my name is Phil."}]},
        {"role": "model", "parts": [{"text": "Hi Phil, how can I help you?"}]},
        {"role": "user", "parts": [{"text": "What is my name?"}]}
    ]
}'

# Response
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "Your name is Phil."
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP",
      "index": 0
    }
  ]
}

Nachher (Interactions API)

Die Interactions API verwaltet den Status auf dem Server. Sie setzen eine Unterhaltung fort, indem Sie auf previous_interaction_id verweisen.

Python

from google import genai

client = genai.Client()

# First turn
interaction1 = client.interactions.create(
    model="gemini-3-flash-preview", input="Hi, my name is Phil."
)
print(interaction1.steps[-1].content[0].text)

# Second turn (passing previous_interaction_id)
interaction2 = client.interactions.create(
    model="gemini-3-flash-preview",
    previous_interaction_id=interaction1.id,
    input="What is my name?",
)
print(interaction2.steps[-1].content[0].text)

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

// First turn
let interaction = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: 'Hi, my name is Phil.'
});
console.log(interaction.steps.at(-1).content[0].text);

// Second turn
interaction = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    previous_interaction_id: interaction.id,
    input: 'What is my name?'
});
console.log(interaction.steps.at(-1).content[0].text);

REST

# First Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "input": "Hi, my name is Phil."
}'

# Second Request (using ID from first response)
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "previous_interaction_id": "int_123",
    "input": "What is my name?"
}'

# Response to Second Request
{
  "id": "int_123",
  "steps": [
    {
      "type": "user_input",
      "status": "done",
      "content": [{ "type": "text", "text": "Hi, my name is Phil." }]
    },
    {
      "type": "model_output",
      "status": "done",
      "content": [{ "type": "text", "text": "Hello Phil! How can I help you today?" }]
    },
    {
      "type": "user_input",
      "status": "done",
      "content": [{ "type": "text", "text": "What is my name?" }]
    },
    {
      "type": "model_output",
      "status": "done",
      "content": [{ "type": "text", "text": "Your name is Phil." }]
    }
  ]
}

Multimodale Eingaben

Beide APIs unterstützen multimodale Eingaben (Text, Bilder, Videos usw.).

Vorher (generateContent)

In generateContent übergeben Sie eine Liste von parts im contents-Array. Die Antwort gibt die Ausgabe in den parts des ersten Kandidaten zurück.

Python

from google import genai
from google.genai import types

client = genai.Client()

with open("sample.jpg", "rb") as f:
    image_bytes = f.read()

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents=[
        types.Part.from_bytes(data=image_bytes, mime_type="image/jpeg"),
        "Describe this image.",
    ],
)
print(response.text)

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [{
        "parts": [
            {
                "inlineData": {
                    "mimeType": "image/jpeg",
                    "data": "..."
                }
            },
            {
                "text": "Describe this image."
            }
        ]
    }]
}'

# Response
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "This is a picture of a beautiful sunset."
          }
        ],
        "role": "model"
      }
    }
  ]
}

Nachher (Interactions API)

In der Interactions API übergeben Sie ein Array an das Feld input. Sie rufen die Ausgabedaten ab, indem Sie den Schritt model_output in der Zeitachse suchen.

Python

from google import genai

client = genai.Client()

# Assuming you have an image file
with open("sample.jpg", "rb") as f:
    image_bytes = f.read()

interaction = client.interactions.create(
    model="gemini-3-flash-preview",
    input=[
        {
            "type": "image",
            "mime_type": "image/jpeg",
            "data": image_bytes,
        },
        {"type": "text", "text": "Describe this image."},
    ],
)
for step in interaction.steps:
    if step.type == "model_output":
        print(step.content[0].text)

JavaScript

import { GoogleGenAI } from '@google/genai';
import * as fs from 'fs';

const client = new GoogleGenAI({});

const imageBytes = fs.readFileSync('sample.jpg').toString('base64');

const interaction = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: [
        {
            type: 'image',
            mime_type: 'image/jpeg',
            data: imageBytes
        },
        {
            type: 'text',
            text: 'Describe this image.'
        }
    ]
});
for (const step of interaction.steps) {
    if (step.type === 'model_output') {
        console.log(step.content[0].text);
    }
}

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "input": [
        {
            "type": "image",
            "mime_type": "image/jpeg",
            "data": "..."
        },
        {
            "type": "text",
            "text": "Describe this image."
        }
    ]
}'

# Response
{
  "id": "int_multimodal",
  "steps": [
    {
      "type": "user_input",
      "status": "done",
      "content": [
        {
          "type": "image",
          "mime_type": "image/jpeg",
          "data": "..."
        },
        {
          "type": "text",
          "text": "Describe this image."
        }
      ]
    },
    {
      "type": "model_output",
      "status": "done",
      "content": [
        {
          "type": "text",
          "text": "This is a picture of a beautiful sunset over the mountains."
        }
      ]
    }
  ]
}

Strukturierte Ausgabe

Wenn das Modell JSON zurückgeben soll, das einem bestimmten Schema entspricht, konfigurieren Sie das Antwortformat.

Vorher (generateContent)

In generateContent konfigurieren Sie das Ausgabeformat mit dem Feld response_format, das im Objekt generationConfig verschachtelt ist.

Python

from google import genai
from google.genai import types
from pydantic import BaseModel

client = genai.Client()

class Recipe(BaseModel):
    recipe_name: str
    ingredients: list[str]

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Give me a recipe for chocolate chip cookies.",
    config=types.GenerateContentConfig(
        response_format=[
            {
                "type": "text",
                "mime_type": "application/json",
                "schema": Recipe,
            }
        ]
    ),
)
print(response.text)

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [{
        "parts": [{
            "text": "Give me a recipe for chocolate chip cookies."
        }]
    }],
    "generationConfig": {
        "responseFormat": [
            {
                "type": "text",
                "mimeType": "application/json",
                "schema": {
                    "type": "OBJECT",
                    "properties": {
                        "recipe_name": { "type": "STRING" },
                        "ingredients": {
                            "type": "ARRAY",
                            "items": { "type": "STRING" }
                        }
                    },
                    "required": ["recipe_name", "ingredients"]
                }
            }
        ]
    }
}'

# Response
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "{\n  \"recipe_name\": \"Chocolate Chip Cookies\",\n  \"ingredients\": [\n    \"1 cup butter\",\n    \"1 cup sugar\",\n    \"2 cups flour\",\n    \"1 cup chocolate chips\"\n  ]\n}"
          }
        ],
        "role": "model"
      }
    }
  ]
}

Nachher (Interactions API)

In der Interactions API werden die Steuerelemente für das Ausgabeformat in ein response_format-Array auf oberster Ebene verschoben.

Python

from google import genai
from pydantic import BaseModel

client = genai.Client()

class Recipe(BaseModel):
    recipe_name: str
    ingredients: list[str]

interaction = client.interactions.create(
    model="gemini-3-flash-preview",
    input="Give me a recipe for chocolate chip cookies.",
    response_format=[
        {
            "type": "text",
            "mime_type": "application/json",
            "schema": Recipe,
        }
    ],
)

for step in interaction.steps:
    if step.type == "model_output":
        print(step.content[0].text)

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const interaction = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: 'Give me a recipe for chocolate chip cookies.',
    response_format: [
        {
            type: 'text',
            mime_type: 'application/json',
            schema: {
                type: 'object',
                properties: {
                    recipe_name: { type: 'string' },
                    ingredients: {
                        type: 'array',
                        items: { type: 'string' }
                    }
                },
                required: ['recipe_name', 'ingredients']
            }
        }
    ]
});
for (const step of interaction.steps) {
    if (step.type === 'model_output') {
        console.log(step.content[0].text);
    }
}

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "input": "Give me a recipe for chocolate chip cookies.",
    "response_format": [
        {
            "type": "text",
            "mime_type": "application/json",
            "schema": {
                "type": "OBJECT",
                "properties": {
                    "recipe_name": { "type": "STRING" },
                    "ingredients": {
                        "type": "ARRAY",
                        "items": { "type": "STRING" }
                    }
                },
                "required": ["recipe_name", "ingredients"]
            }
        }
    ]
}'

# Response
{
  "id": "int_structured",
  "steps": [
    {
      "type": "user_input",
      "status": "done",
      "content": [{ "type": "text", "text": "Give me a recipe for chocolate chip cookies." }]
    },
    {
      "type": "model_output",
      "status": "done",
      "content": [
        {
          "type": "text",
          "text": "{\n  \"recipe_name\": \"Chocolate Chip Cookies\",\n  \"ingredients\": [\n    \"1 cup butter\",\n    \"1 cup sugar\",\n    \"2 cups flour\",\n    \"1 cup chocolate chips\"\n  ]\n}"
        }
      ]
    }
  ]
}

Multimodale Generierung

Wenn Sie Inhalte in anderen Modalitäten als Text (z. B. Bilder oder Audio) generieren, besteht der Hauptunterschied darin, wie die Antwort die generierten Medien strukturiert.

Vorher (generateContent)

In generateContent gibt die Antwort generierte Medien direkt in den parts des Kandidaten zurück, in der Regel als Base64-Daten in inlineData.

# Response structure concept
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "Here is your generated image:"
          },
          {
            "inlineData": {
              "mimeType": "image/jpeg",
              "data": "...base64..."
            }
          }
        ]
      }
    }
  ]
}

Nachher (Interactions API)

In der Interactions API werden generierte Medien als separate Elemente im content-Array eines model_output-Schritts in der Zeitachse angezeigt, wodurch der chronologische Ablauf der Interaktion beibehalten wird.

# Response structure concept
{
  "id": "int_123",
  "steps": [
    {
      "type": "model_output",
      "status": "done",
      "content": [
        {
          "type": "text",
          "text": "Here is your generated image:"
        },
        {
          "type": "image",
          "mime_type": "image/jpeg",
          "data": "...base64..." // Or a reference URL in future
        }
      ]
    }
  ]
}

Dadurch bleibt das Parsen der Antwort mit der Verarbeitung von Eingaben und Textausgaben konsistent. Alles ist ein Schritt in der Zeitachse.

Serverseitige Tools

Gemini unterstützt integrierte serverseitige Tools wie die Google Suche. Der Hauptunterschied besteht darin, wie die Antwort die Toolausführung darstellt.

Vorher (generateContent)

In generateContent sind serverseitige Tools weitgehend undurchsichtig. Sie aktivieren das Tool und erhalten eine endgültige Antwort mit einem separaten groundingMetadata-Objekt. Wichtig ist, dass Zitate nicht inline sind. groundingSupports verwendet Zeichenindizes, um Textsegmente in groundingChunks wieder Webquellen zuzuordnen.

Python

from google import genai
from google.genai import types

client = genai.Client()

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Who won Euro 2024?",
    config=types.GenerateContentConfig(
        tools=[{"google_search": {}}]
    ),
)

# Access search entry point (widget) and citations
metadata = response.candidates[0].grounding_metadata
if metadata.search_entry_point:
    print(f"Search Entry Point: {metadata.search_entry_point.rendered_content}")

for support in metadata.grounding_supports:
    print(f"Citation: {support.segment.text}")

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const response = await client.models.generateContent({
    model: 'gemini-2.5-flash',
    contents: 'Who won Euro 2024?',
    config: {
        tools: [{ google_search: {} }]
    }
});

const metadata = response.candidates[0].groundingMetadata;
if (metadata.searchEntryPoint) {
    console.log(`Search Entry Point: ${metadata.searchEntryPoint.renderedContent}`);
}
for (const support of metadata.groundingSupports) {
    console.log(`Citation: ${support.segment.text}`);
}

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [{
        "parts": [{
            "text": "Who won Euro 2024?"
        }]
    }],
    "tools": [{
        "googleSearchRetrieval": {}
    }]
}'

# Response
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "Spain won Euro 2024, defeating England 2-1 in the final. This victory marks Spain's record fourth European Championship title."
          }
        ],
        "role": "model"
      },
      "groundingMetadata": {
        "webSearchQueries": [
          "UEFA Euro 2024 winner",
          "who won euro 2024"
        ],
        "searchEntryPoint": {
          "renderedContent": "<!-- HTML and CSS for the search widget -->"
        },
        "groundingChunks": [
          {"web": {"uri": "https://vertexaisearch.cloud.google.com.....", "title": "aljazeera.com"}},
          {"web": {"uri": "https://vertexaisearch.cloud.google.com.....", "title": "uefa.com"}}
        ],
        "groundingSupports": [
          {
            "segment": {"startIndex": 0, "endIndex": 85, "text": "Spain won Euro 2024, defeatin..."},
            "groundingChunkIndices": [0]
          },
          {
            "segment": {"startIndex": 86, "endIndex": 210, "text": "This victory marks Spain's..."},
            "groundingChunkIndices": [0, 1]
          }
        ]
      }
    }
  ]
}

Nachher (Interactions API)

In der Interactions API bieten serverseitige Tools vollständige Transparenz für die Zeitachse. Die API erfasst den Aufruf und das Ergebnis als separate Ausführung steps (google_search_call und google_search_result) und zeigt genau, welche Daten das Modell abgerufen hat.

Außerdem gibt die API Zitate inline zurück. Anstatt Indizes aus einem separaten Metadatenobjekt zuzuordnen, enthält das Textelement im Schritt model_output ein eigenes annotations-Array, das direkt mit der Quelle verknüpft ist.

Python

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-3-flash-preview",
    input="Who won Euro 2024?",
    tools=[{"type": "google_search"}],
)

for step in interaction.steps:
    if step.type == "google_search_result":
        print(f"Search Suggestions: {step.search_suggestions}")
    elif step.type == "model_output":
        print(f"Answer: {step.content[0].text}")
        if step.content[0].annotations:
            for anno in step.content[0].annotations:
                print(f"Citation: {anno.title} ({anno.uri})")

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const interaction = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: 'Who won Euro 2024?',
    tools: [{ type: 'google_search' }]
});

for (const step of interaction.steps) {
    if (step.type === 'google_search_result') {
        console.log(`Search Suggestions: ${step.search_suggestions}`);
    } else if (step.type === 'model_output') {
        console.log(`Answer: ${step.content[0].text}`);
        if (step.content[0].annotations) {
            for (const anno of step.content[0].annotations) {
                console.log(`Citation: ${anno.title} (${anno.uri})`);
            }
        }
    }
}

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "input": "Who won Euro 2024?",
    "tools": [{"type": "google_search"}]
}'

# Response (showing grounding)
{
  "id": "int_grounded",
  "steps": [
    {
      "type": "user_input",
      "status": "done",
      "content": [{ "type": "text", "text": "Who won Euro 2024?" }]
    },
    {
      "type": "google_search_call",
      "status": "done",
      "content": [{ "type": "text", "text": "UEFA Euro 2024 winner" }]
    },
    {
      "type": "google_search_result",
      "status": "done",
      "content": [
        {
          "type": "text",
          "text": "Spain won Euro 2024..." 
        }
      ]
    },
    {
      "type": "model_output",
      "status": "done",
      "content": [
        {
          "type": "text",
          "text": "Spain won Euro 2024, defeating England 2-1.",
          "annotations": [
            {
              "start_index": 0,
              "end_index": 42,
              "uri": "https://vertexaisearch...",
              "title": "aljazeera.com"
            }
          ]
        }
      ]
    }
  ]
}

Funktionsaufrufe

Die Struktur von Funktionsaufrufen und ‑ergebnissen wurde ebenfalls an das Schema „Steps“ angepasst.

Vorher (generateContent)

In generateContent gibt die Antwort Funktionsaufrufe innerhalb der Kandidaten zurück.

Python

from google import genai
from google.genai import types

client = genai.Client()

# Step 1: Send prompt with tools
response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="What's the weather in Boston?",
    config=types.GenerateContentConfig(tools=[weather_tool]),
)

# Assume model returned function_call
function_call = response.candidates[0].content.parts[0].function_call
print(f"Requested tool: {function_call.name}")

# Step 2: Execute local function and send result back
result = "52°F and rain"

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents=[
        types.Content(
            role="user",
            parts=[
                types.Part.from_text(text="What's the weather in Boston?")
            ],
        ),
        response.candidates[0].content,  # Model turn with function call
        types.Content(
            role="user",
            parts=[
                types.Part.from_function_response(
                    name=function_call.name,
                    response={"result": result},
                )
            ],
        ),
    ],
    config=types.GenerateContentConfig(tools=[weather_tool]),
)
print(response.text)

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

// Step 1: Send prompt with tools
let response = await client.models.generateContent({
    model: 'gemini-2.5-flash',
    contents: "What's the weather in Boston?",
    config: { tools: [weatherTool] }
});

const functionCall = response.candidates[0].content.parts[0].functionCall;
console.log(`Requested tool: ${functionCall.name}`);

// Step 2: Execute local function and send result back
const result = "52°F and rain";

response = await client.models.generateContent({
    model: 'gemini-2.5-flash',
    contents: [
        { role: 'user', parts: [{ text: "What's the weather in Boston?" }] },
        response.candidates[0].content, // Model turn
        {
            role: 'user',
            parts: [{
                functionResponse: {
                    name: functionCall.name,
                    response: { result: result }
                }
            }]
        }
    ],
    config: { tools: [weatherTool] }
});
console.log(response.text);

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [{
        "parts": [{
            "text": "What is the weather like in Boston, MA?"
        }]
    }],
    "tools": [{
        "functionDeclarations": [{
            "name": "get_weather",
            "description": "Get the current weather",
            "parameters": {
                "type": "OBJECT",
                "properties": {
                    "location": {"type": "STRING"}
                },
                "required": ["location"]
            }
        }]
    }]
}'

# Response
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "functionCall": {
              "name": "get_weather",
              "args": { "location": "Boston, MA" }
            }
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP",
      "index": 0
    }
  ]
}

Nachher (Interactions API)

Toolaufrufe und ‑ergebnisse sind jetzt separate Schritte in der Zeitachse.

Python

from google import genai

client = genai.Client()

weather_tool = {
    "type": "function",
    "name": "get_weather",
    "description": "Gets weather",
    "parameters": {
        "type": "object",
        "properties": {
            "location": {"type": "string"}
        },
    },
}

interaction = client.interactions.create(
    model="gemini-3-flash-preview",
    input="What's the weather in Boston?",
    tools=[weather_tool],
)

# Check if the model requested a tool call
for step in interaction.steps:
    if step.type == "function_call":
        print(f"Executing {step.name} for {step.arguments}")

        # Execute your local function here...
        result = "52°F and rain"

        # Submit the result back as a step
        interaction = client.interactions.create(
            model="gemini-3-flash-preview",
            previous_interaction_id=interaction.id,
            input=[
                {
                    "type": "function_result",
                    "call_id": step.id,
                    "name": step.name,
                    "result": [{"type": "text", "text": result}],
                }
            ],
        )
        # Inspect steps for final response
        for s in interaction.steps:
            if s.type == "model_output":
                print(s.content[0].text)

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const weatherTool = {
    type: "function",
    name: "get_weather",
    description: "Get weather for a location",
    parameters: {
        type: "object",
        properties: {
            location: { type: "string" }
        },
        required: ["location"]
    }
};

const interaction = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: "What's the weather in Boston?",
    tools: [weatherTool]
});

// Check if the model requested a tool call
for (const step of interaction.steps) {
    if (step.type === 'function_call') {
        console.log(`Executing ${step.name} for ${JSON.stringify(step.arguments)}`);

        const result = "52°F and rain";

        // Submit the result back as a step
        const nextInteraction = await client.interactions.create({
            model: 'gemini-3-flash-preview',
            previous_interaction_id: interaction.id,
            input: [
                {
                    type: 'function_result',
                    call_id: step.id,
                    name: step.name,
                    result: [{ type: 'text', text: result }]
                }
            ]
        });

        // Inspect steps for final response
        for (const s of nextInteraction.steps) {
            if (s.type === 'model_output') {
                console.log(s.content[0].text);
            }
        }
    }
}

REST

# Initial Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "input": "What's the weather in Boston?",
    "tools": [{
        "type": "function",
        "name": "get_weather",
        "description": "Get weather for a location",
        "parameters": {
            "type": "object",
            "properties": {
                "location": { "type": "string" }
            },
            "required": ["location"]
        }
    }]
}'

# Response (requires action)
{
  "id": "int_001",
  "status": "requires_action",
  "steps": [
    {
      "type": "user_input",
      "status": "done",
      "content": [
        { "type": "text", "text": "What's the weather in Boston?" }
      ]
    },
    {
      "type": "function_call",
      "status": "waiting",
      "id": "fc_1",
      "name": "get_weather",
      "arguments": { "location": "Boston, MA" }
    }
  ]
}

# Submit Tool Result Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "previous_interaction_id": "int_001",
    "input": {
        "type": "function_result",
        "call_id": "fc_1",
        "name": "get_weather",
        "result": [
            { "type": "text", "text": "52°F with rain" }
        ]
    }
}'

# Final Response
{
  "id": "int_002",
  "status": "completed",
  "steps": [
    {
      "type": "function_result",
      "call_id": "fc_1",
      "name": "get_weather",
      "result": [
        { "type": "text", "text": "52°F with rain" }
      ]
    },
    {
      "type": "model_output",
      "status": "done",
      "content": [
        { "type": "text", "text": "It's 52°F with rain in Boston." }
      ]
    }
  ]
}

Streaming

Ein wichtiger Unterschied beim Streaming besteht darin, dass die Interactions API denselben Endpunkt mit "stream": true im Anfragetext verwendet, während für die generateContent API ein dedizierter Endpunkt (:streamGenerateContent) aufgerufen werden musste.

Außerdem verwenden Streamingereignisse jetzt spezielle Typen, um den Lebenszyklus der Interaktion zu beobachten und Ausführungsschritte entlang der Zeitachse zu verfolgen.

Vorher (generateContentStream)

Mit generateContent nutzen Sie einen Stream von Antwort-Chunks.

Python

response = client.models.generate_content_stream(
    model="gemini-2.5-flash", contents="Tell me a story"
)
for chunk in response:
    print(chunk.text, end="")

JavaScript

const responseStream = await client.models.generateContentStream({
    model: 'gemini-2.5-flash',
    contents: 'Tell me a story',
});
for await (const chunk of responseStream) {
    process.stdout.write(chunk.text);
}

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:streamGenerateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [{
        "parts": [{
            "text": "Tell me a story"
        }]
    }]
}'

# Response stream
event: content.start
data: {"event_type": "content.start", "index": 0, "content": {"type": "thought"}}
event: content.delta
data: {"event_type": "content.delta", "index": 0, "delta": {"type": "thought_summary", "text": "User wants an explanation."}}
event: content.stop
data: {"event_type": "content.stop", "index": 0}
event: content.start
data: {"event_type": "content.start", "index": 1, "content": {"type": "text"}}
event: content.delta
data: {"event_type": "content.delta", "index": 1, "delta": {"type": "text", "text": "Hello"}}
event: content.stop
data: {"event_type": "content.stop", "index": 1}

Nachher (Interactions API)

In der Interactions API verwendet Streaming vom Server gesendete Ereignisse (SSE, Server-Sent Events) und spezielle Deltas, um Ausführungsschritte darzustellen.

Python

from google import genai

client = genai.Client()

stream = client.interactions.create(
    model="gemini-3-flash-preview",
    input="Tell me a story",
    stream=True,
)

for event in stream:
    if event.event_type == "step.delta":
        if event.delta.type == "text":
            print(event.delta.text, end="", flush=True)
    elif event.event_type == "interaction.complete":
        print(f"\n\n--- Stream Finished ---")

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const stream = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: 'Tell me a story',
    stream: true,
});

for await (const event of stream) {
    if (event.event_type === 'step.delta') {
        if (event.delta.type === 'text' && 'text' in event.delta) {
            process.stdout.write(event.delta.text);
        }
    } else if (event.event_type === 'interaction.complete') {
        console.log('\n\n--- Stream Finished ---');
    }
}

REST

# Example SSE stream output event: interaction.created data: {"type": "interaction.created", "interaction": {"id": "int_xyz", "status": "created"}} event: interaction.status_update data: {"type": "interaction.status_update", "status": "in_progress"}} event: step.start data: {"type": "step.start", "index": 0, "step": {"type": "thought"}} event: step.delta data: {"type": "step.delta", "index": 0, "delta": {"type": "thought", "text": "User wants an explanation."}} event: step.stop data: {"type": "step.stop", "index": 0, "status": "done"}} event: step.start data: {"type": "step.start", "index": 1, "step": {"type": "model_output"}} event: step.delta data: {"type": "step.delta", "index": 1, "delta": {"type": "text", "text": "Hello"}} event: step.stop data: {"type": "step.stop", "index": 1, "status": "done"}} event: interaction.complete data: {"type": "interaction.complete", "interaction": {"id": "int_xyz", "status": "completed", "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}}} ```

Streaming-Tools und Funktionsaufrufe

Das Verhalten von Tools im Stream hat sich von generateContent deutlich geändert, um eine detailliertere Steuerung und Sichtbarkeit zu ermöglichen.

Vorher (generateContent)

Mit generateContent kamen Streaming-Funktionsaufrufe vollständig in einem einzigen Chunk an. Sie konnten die Argumente nicht in Echtzeit generieren sehen, daher wurde vom Handler einfach nach einem vollständigen functionCall-Objekt gesucht.

Python

from google import genai
from google.genai import types

client = genai.Client()

stream = client.models.generate_content_stream(
    model="gemini-2.5-flash",
    contents="What's the weather in Boston?",
    config=types.GenerateContentConfig(tools=[weather_tool]),
)

for chunk in stream:
    # Function calls arrived complete — no partial arguments
    if chunk.candidates[0].content.parts[0].function_call:
        fc = chunk.candidates[0].content.parts[0].function_call
        print(f"Call: {fc.name}({fc.args})")
    elif chunk.text:
        print(chunk.text, end="")

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const stream = await client.models.generateContentStream({
    model: 'gemini-2.5-flash',
    contents: "What's the weather in Boston?",
    config: { tools: [weatherTool] }
});

for await (const chunk of stream) {
    // Function calls arrived complete — no partial arguments
    const part = chunk.candidates[0].content.parts[0];
    if (part.functionCall) {
        console.log(`Call: ${part.functionCall.name}(${JSON.stringify(part.functionCall.args)})`);
    } else if (part.text) {
        process.stdout.write(part.text);
    }
}

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:streamGenerateContent" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "contents": [{"parts": [{"text": "What'\''s the weather in Boston?"}]}],
    "tools": [{"functionDeclarations": [{"name": "get_weather", "parameters": {"type": "OBJECT", "properties": {"location": {"type": "STRING"}}}}]}]
}'

# Response stream — function call arrives complete in one chunk
{"candidates": [{"content": {"parts": [{"functionCall": {"name": "get_weather", "args": {"location": "Boston, MA"}}}]}}]}

Nachher (Interactions API)

Die Interactions API streamt Funktionsaufrufargumente zeichenweise als arguments-Ereignisse. Der gesamte Toollebenszyklus – Gedanken, Aufruf, Ergebnis und Ausgabe – wird als eine Reihe separater Schritte dargestellt.

Python

from google import genai

client = genai.Client()

stream = client.interactions.create(
    model="gemini-3-flash-preview",
    input="What's the weather in Boston?",
    tools=[get_weather_tool],
    stream=True,
)

for event in stream:
    if event.event_type == "step.start":
        if event.step.type == "function_call":
            print(f"Calling: {event.step.name}")
    elif event.event_type == "step.delta":
        if event.delta.type == "arguments":
            print(f"  args: {event.delta.partial_arguments}")
        elif event.delta.type == "text":
            print(event.delta.text, end="")
    elif event.event_type == "interaction.complete":
        print("\n--- Done ---")

JavaScript

import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});

const stream = await client.interactions.create({
    model: 'gemini-3-flash-preview',
    input: "What's the weather in Boston?",
    tools: [getWeatherTool],
    stream: true,
});

for await (const event of stream) {
    if (event.event_type === 'step.start') {
        if (event.step.type === 'function_call') {
            console.log(`Calling: ${event.step.name}`);
        }
    } else if (event.event_type === 'step.delta') {
        if (event.delta.type === 'arguments') {
            console.log(`  args: ${event.delta.partial_arguments}`);
        } else if (event.delta.type === 'text') {
            process.stdout.write(event.delta.text);
        }
    } else if (event.event_type === 'interaction.complete') {
        console.log('\n--- Done ---');
    }
}

REST

# Request
curl -X POST "https://generativelanguage.googleapis.com/v1beta2/interactions" \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
    "model": "gemini-3-flash-preview",
    "input": "What'\''s the weather in Boston?",
    "tools": [{"type": "function", "name": "get_weather", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}}}],
    "stream": true
}'

# Response stream
// Interaction created
event: interaction.created
data: {"type": "interaction.created", "interaction": {"id": "int_xyz", "status": "created"}}

event: interaction.status_update
data: {"type": "interaction.status_update", "status": "in_progress"}

// ── Step 0: Thought ──────────────────────────────────
event: step.start
data: {"type": "step.start", "index": 0, "step": {"type": "thought"}}

event: step.delta
data: {"type": "step.delta", "index": 0, "delta": {"type": "thought", "text": "The user wants weather data for Boston. I'll call the get_weather tool."}}

event: step.stop
data: {"type": "step.stop", "index": 0, "status": "done"}

// ── Step 1: Function Call (arguments streamed) ───────
event: step.start
data: {"type": "step.start", "index": 1, "step": {"type": "function_call", "id": "fc_1", "name": "get_weather"}}

event: step.delta
data: {"type": "step.delta", "index": 1, "delta": {"type": "arguments", "partial_arguments": "{\"location\": \"Boston, MA\"}"}}

event: step.stop
data: {"type": "step.stop", "index": 1, "status": "waiting"}

// The interaction pauses — the model needs the tool result before continuing.
event: interaction.status_update
data: {"type": "interaction.status_update", "status": "requires_action"}

// ── (Client submits the tool result) ──────────────────
// The client calls interactions.create with the function_result as input
// and the previous interaction's ID, then resumes consuming the stream.

event: interaction.status_update
data: {"type": "interaction.status_update", "status": "in_progress"}

// ── Step 2: Function Result (echoed back, no deltas) ─
event: step.start
data: {"type": "step.start", "index": 2, "step": {"type": "function_result", "call_id": "fc_1", "name": "get_weather", "result": [{"type": "text", "text": "52°F, rain"}]}}

event: step.stop
data: {"type": "step.stop", "index": 2, "status": "done"}

// ── Step 3: Thought ──────────────────────────────────
event: step.start
data: {"type": "step.start", "index": 3, "step": {"type": "thought"}}

event: step.delta
data: {"type": "step.delta", "index": 3, "delta": {"type": "thought", "text": "Got weather data. Composing the final response."}}

event: step.stop
data: {"type": "step.stop", "index": 3, "status": "done"}

// ── Step 4: Model Output (text streamed) ─────────────
event: step.start
data: {"type": "step.start", "index": 4, "step": {"type": "model_output"}}

event: step.delta
data: {"type": "step.delta", "index": 4, "delta": {"type": "text", "text": "It's currently 52°F and rainy in Boston."}}

event: step.stop
data: {"type": "step.stop", "index": 4, "status": "done"}

// ── Interaction complete ─────────────────────────────
event: interaction.complete
data: {"type": "interaction.complete", "interaction": {"id": "int_xyz", "status": "completed", "usage": {"prompt_tokens": 256, "completion_tokens": 128, "total_tokens": 384}}}

Serverseitige Tools im Stream

Serverseitige Tools wie die Google Suche verhalten sich im Stream anders als Funktionsaufrufe. Ihr Aufruf und Ergebnis kommen vollständig im step.start-Ereignis ohne Deltas an. Auf step.start folgt sofort step.stop:

// Server-side tool call — payload arrives complete in step.start
event: step.start
data: {"type": "step.start", "index": 4, "step": {"type": "google_search_call", "id": "gs_2", "query": "Alphabet Q4 2025 earnings"}}

event: step.stop
data: {"type": "step.stop", "index": 4, "status": "done"}

// Server-side tool result — also complete in step.start
event: step.start
data: {"type": "step.start", "index": 5, "step": {"type": "google_search_result", "call_id": "gs_2", "rendered_content": "<div>Alphabet Q4 2025 Revenue: $105.6B</div>", "signature": "abc123..."}}

event: step.stop
data: {"type": "step.stop", "index": 5, "status": "done"}