feat: support structured outputs in MetaLlamaChatGenerator

integrations/meta_llama/examples/chat_generator_with_structured_output.py

@@ -0,0 +1,35 @@
+# SPDX-FileCopyrightText: 2024-present deepset GmbH <[email protected]>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+
+# This example demonstrates how to use the MetaLlamaChatGenerator component
+# with structured outputs.
+# To run this example, you will need to
+# set `LLAMA_API_KEY` environment variable
+
+from haystack.dataclasses import ChatMessage
+from pydantic import BaseModel
+
+from haystack_integrations.components.generators.meta_llama import MetaLlamaChatGenerator
+
+
+class NobelPrizeInfo(BaseModel):
+    recipient_name: str
+    award_year: int
+    category: str
+    achievement_description: str
+    nationality: str
+
+
+chat_messages = [
+    ChatMessage.from_user(
+        "In 2021, American scientist David Julius received the Nobel Prize in"
+        " Physiology or Medicine for his groundbreaking discoveries on how the human body"
+        " senses temperature and touch."
+    )
+]
+component = MetaLlamaChatGenerator(generation_kwargs={"response_format": NobelPrizeInfo})
+results = component.run(chat_messages)
+
+# print(results)

integrations/meta_llama/pyproject.toml

@@ -23,7 +23,7 @@ classifiers = [
   "Programming Language :: Python :: Implementation :: CPython",
   "Programming Language :: Python :: Implementation :: PyPy",
 ]
-dependencies = ["haystack-ai>=2.13.2"]
+dependencies = ["haystack-ai>=2.19.0"]
 
 [project.urls]
 Documentation = "https://github.com/deepset-ai/haystack-core-integrations/tree/main/integrations/meta_llama#readme"

integrations/meta_llama/src/haystack_integrations/components/generators/meta_llama/chat/chat_generator.py

@@ -3,14 +3,16 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, Optional
 
 from haystack import component, default_to_dict, logging
 from haystack.components.generators.chat import OpenAIChatGenerator
 from haystack.dataclasses import ChatMessage, StreamingCallbackT
-from haystack.tools import Tool, Toolset
+from haystack.tools import ToolsType, serialize_tools_or_toolset
 from haystack.utils import serialize_callable
 from haystack.utils.auth import Secret
+from openai.lib._pydantic import to_strict_json_schema
+from pydantic import BaseModel
 
 logger = logging.getLogger(__name__)
 
@@ -60,7 +62,7 @@ def __init__(
         streaming_callback: Optional[StreamingCallbackT] = None,
         api_base_url: Optional[str] = "https://api.llama.com/compat/v1/",
         generation_kwargs: Optional[Dict[str, Any]] = None,
-        tools: Optional[Union[List[Tool], Toolset]] = None,
+        tools: Optional[ToolsType] = None,
     ):
         """
         Creates an instance of LlamaChatGenerator. Unless specified otherwise in the `model`, this is for Llama's
@@ -91,6 +93,12 @@ def __init__(
                 events as they become available, with the stream terminated by a data: [DONE] message.
             - `safe_prompt`: Whether to inject a safety prompt before all conversations.
             - `random_seed`: The seed to use for random sampling.
+            - `response_format`: A JSON schema or a Pydantic model that enforces the structure of the model's response.
+                If provided, the output will always be validated against this
+                format (unless the model returns a tool call).
+                For details, see the [OpenAI Structured Outputs documentation](https://platform.openai.com/docs/guides/structured-outputs).
+                For structured outputs with streaming, the `response_format` must be a JSON
+                schema and not a Pydantic model.
         :param tools:
             A list of tools for which the model can prepare calls.
         """
@@ -110,7 +118,7 @@ def _prepare_api_call(
         messages: list[ChatMessage],
         streaming_callback: Optional[StreamingCallbackT] = None,
         generation_kwargs: Optional[dict[str, Any]] = None,
-        tools: Optional[Union[list[Tool], Toolset]] = None,
+        tools: Optional[ToolsType] = None,
         tools_strict: Optional[bool] = None,
     ) -> dict[str, Any]:
         api_args = super(MetaLlamaChatGenerator, self)._prepare_api_call(  # noqa: UP008
@@ -133,13 +141,29 @@ def to_dict(self) -> Dict[str, Any]:
             The serialized component as a dictionary.
         """
         callback_name = serialize_callable(self.streaming_callback) if self.streaming_callback else None
+        generation_kwargs = self.generation_kwargs.copy()
+        response_format = generation_kwargs.get("response_format")
+
+        # If the response format is a Pydantic model, it's converted to openai's json schema format
+        # If it's already a json schema, it's left as is
+        if response_format and isinstance(response_format, type) and issubclass(response_format, BaseModel):
+            json_schema = {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": response_format.__name__,
+                    "strict": True,
+                    "schema": to_strict_json_schema(response_format),
+                },
+            }
+
+            generation_kwargs["response_format"] = json_schema
 
         return default_to_dict(
             self,
             model=self.model,
             streaming_callback=callback_name,
             api_base_url=self.api_base_url,
-            generation_kwargs=self.generation_kwargs,
+            generation_kwargs=generation_kwargs,
             api_key=self.api_key.to_dict(),
-            tools=[tool.to_dict() for tool in self.tools] if self.tools else None,
+            tools=serialize_tools_or_toolset(self.tools),
         )

integrations/meta_llama/tests/test_llama_chat_generator.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates
 
+import json
 import os
 from datetime import datetime
 from unittest.mock import patch
@@ -15,6 +16,7 @@
 from openai import OpenAIError
 from openai.types.chat import ChatCompletion, ChatCompletionMessage
 from openai.types.chat.chat_completion import Choice
+from pydantic import BaseModel
 
 from haystack_integrations.components.generators.meta_llama.chat.chat_generator import (
     MetaLlamaChatGenerator,
@@ -134,12 +136,44 @@ def test_to_dict_default(self, monkeypatch):
 
     def test_to_dict_with_parameters(self, monkeypatch):
         monkeypatch.setenv("ENV_VAR", "test-api-key")
+
+        class NobelPrizeInfo(BaseModel):
+            recipient_name: str
+            award_year: int
+
+        schema = {
+            "json_schema": {
+                "name": "NobelPrizeInfo",
+                "schema": {
+                    "additionalProperties": False,
+                    "properties": {
+                        "award_year": {
+                            "title": "Award Year",
+                            "type": "integer",
+                        },
+                        "recipient_name": {
+                            "title": "Recipient Name",
+                            "type": "string",
+                        },
+                    },
+                    "required": [
+                        "recipient_name",
+                        "award_year",
+                    ],
+                    "title": "NobelPrizeInfo",
+                    "type": "object",
+                },
+                "strict": True,
+            },
+            "type": "json_schema",
+        }
+
         component = MetaLlamaChatGenerator(
             api_key=Secret.from_env_var("ENV_VAR"),
             model="Llama-4-Scout-17B-16E-Instruct-FP8",
             streaming_callback=print_streaming_chunk,
             api_base_url="test-base-url",
-            generation_kwargs={"max_tokens": 10, "some_test_param": "test-params"},
+            generation_kwargs={"max_tokens": 10, "some_test_param": "test-params", "response_format": NobelPrizeInfo},
         )
         data = component.to_dict()
 
@@ -153,7 +187,7 @@ def test_to_dict_with_parameters(self, monkeypatch):
             "model": "Llama-4-Scout-17B-16E-Instruct-FP8",
             "api_base_url": "test-base-url",
             "streaming_callback": "haystack.components.generators.utils.print_streaming_chunk",
-            "generation_kwargs": {"max_tokens": 10, "some_test_param": "test-params"},
+            "generation_kwargs": {"max_tokens": 10, "some_test_param": "test-params", "response_format": schema},
         }
 
         for key, value in expected_params.items():
@@ -293,6 +327,77 @@ def __call__(self, chunk: StreamingChunk) -> None:
         assert callback.counter > 1
         assert "Paris" in callback.responses
 
+    @pytest.mark.skipif(
+        not os.environ.get("LLAMA_API_KEY", None),
+        reason="Export an env var called LLAMA_API_KEY containing the Llama API key to run this test.",
+    )
+    @pytest.mark.integration
+    def test_live_run_response_format(self):
+        class NobelPrizeInfo(BaseModel):
+            recipient_name: str
+            award_year: int
+            category: str
+            achievement_description: str
+            nationality: str
+
+        chat_messages = [
+            ChatMessage.from_user(
+                "In 2021, American scientist David Julius received the Nobel Prize in"
+                " Physiology or Medicine for his groundbreaking discoveries on how the human body"
+                " senses temperature and touch."
+            )
+        ]
+        component = MetaLlamaChatGenerator(generation_kwargs={"response_format": NobelPrizeInfo})
+        results = component.run(chat_messages)
+        assert isinstance(results, dict)
+        assert "replies" in results
+        assert isinstance(results["replies"], list)
+        assert len(results["replies"]) == 1
+        assert isinstance(results["replies"][0], ChatMessage)
+        message = results["replies"][0]
+        assert isinstance(message.text, str)
+        msg = json.loads(message.text)
+        assert msg["recipient_name"] == "David Julius"
+        assert msg["award_year"] == 2021
+        assert "category" in msg
+        assert "achievement_description" in msg
+        assert msg["nationality"] == "American"
+
+    @pytest.mark.skipif(
+        not os.environ.get("LLAMA_API_KEY", None),
+        reason="Export an env var called LLAMA_API_KEY containing the Llama API key to run this test.",
+    )
+    @pytest.mark.integration
+    def test_live_run_with_response_format_json_schema(self):
+        response_schema = {
+            "type": "json_schema",
+            "json_schema": {
+                "name": "CapitalCity",
+                "strict": True,
+                "schema": {
+                    "title": "CapitalCity",
+                    "type": "object",
+                    "properties": {
+                        "city": {"title": "City", "type": "string"},
+                        "country": {"title": "Country", "type": "string"},
+                    },
+                    "required": ["city", "country"],
+                    "additionalProperties": False,
+                },
+            },
+        }
+
+        chat_messages = [ChatMessage.from_user("What's the capital of France?")]
+        comp = MetaLlamaChatGenerator(generation_kwargs={"response_format": response_schema})
+        results = comp.run(chat_messages)
+        assert len(results["replies"]) == 1
+        message: ChatMessage = results["replies"][0]
+        msg = json.loads(message.text)
+        assert "Paris" in msg["city"]
+        assert isinstance(msg["country"], str)
+        assert "France" in msg["country"]
+        assert message.meta["finish_reason"] == "stop"
+
     @pytest.mark.skipif(
         not os.environ.get("LLAMA_API_KEY", None),
         reason="Export an env var called LLAMA_API_KEY containing the OpenAI API key to run this test.",