Adding support for structured outputs

instructor/client.py

@@ -425,6 +425,7 @@ def from_openai(
             instructor.Mode.FUNCTIONS,
             instructor.Mode.PARALLEL_TOOLS,
             instructor.Mode.MD_JSON,
+            instructor.Mode.STRUCTURED_OUTPUTS,
         }
 
     if isinstance(client, openai.OpenAI):

instructor/function_calls.py

@@ -264,7 +264,7 @@ def from_response(
             Mode.warn_mode_functions_deprecation()
             return cls.parse_functions(completion, validation_context, strict)
 
-        if mode in {Mode.TOOLS, Mode.MISTRAL_TOOLS}:
+        if mode in {Mode.TOOLS, Mode.MISTRAL_TOOLS, Mode.STRUCTURED_OUTPUTS}:
             return cls.parse_tools(completion, validation_context, strict)
 
         if mode in {Mode.JSON, Mode.JSON_SCHEMA, Mode.MD_JSON}:

instructor/mode.py

@@ -19,6 +19,7 @@ class Mode(enum.Enum):
     VERTEXAI_JSON = "vertexai_json"
     GEMINI_JSON = "gemini_json"
     COHERE_JSON_SCHEMA = "json_object"
+    STRUCTURED_OUTPUTS = "structured_output"
 
     @classmethod
     def warn_mode_functions_deprecation(cls):

instructor/process_response.py

@@ -12,6 +12,7 @@
 from instructor.utils import merge_consecutive_messages
 from instructor.validators import AsyncValidationError
 from openai.types.chat import ChatCompletion
+from openai import pydantic_function_tool
 from pydantic import BaseModel, create_model
 import json
 import inspect
@@ -249,6 +250,15 @@ def handle_response_model(
             Mode.warn_mode_functions_deprecation()
             new_kwargs["functions"] = [response_model.openai_schema]
             new_kwargs["function_call"] = {"name": response_model.openai_schema["name"]}
+        elif mode in {Mode.STRUCTURED_OUTPUTS}:
+            response_model_schema = pydantic_function_tool(response_model)
+            response_model_schema["function"]["strict"] = True
+            new_kwargs["tools"] = [response_model_schema]
+
+            new_kwargs["tool_choice"] = {
+                "type": "function",
+                "function": {"name": response_model_schema["function"]["name"]},
+            }
         elif mode in {Mode.TOOLS, Mode.MISTRAL_TOOLS}:
             new_kwargs["tools"] = [
                 {
@@ -263,6 +273,7 @@ def handle_response_model(
                     "type": "function",
                     "function": {"name": response_model.openai_schema["name"]},
                 }
+
         elif mode in {Mode.JSON, Mode.MD_JSON, Mode.JSON_SCHEMA}:
             # If its a JSON Mode we need to massage the prompt a bit
             # in order to get the response we want in a json format

instructor/retry.py

@@ -108,7 +108,7 @@ def reask_messages(response: ChatCompletion, mode: Mode, exception: Exception):
 
     yield dump_message(response.choices[0].message)
     # TODO: Give users more control on configuration
-    if mode == Mode.TOOLS:
+    if mode == Mode.TOOLS or mode == Mode.STRUCTURED_OUTPUTS:
         for tool_call in response.choices[0].message.tool_calls:
             yield {
                 "role": "tool",

instructor/utils.py

@@ -166,6 +166,14 @@ def dump_message(message: ChatCompletionMessage) -> ChatCompletionMessageParam:
         and message.function_call is not None
         and ret["content"]
     ):
+        if not isinstance(ret["content"], str):
+            response_message: str = ""
+            for content_message in ret["content"]:
+                if "text" in content_message:
+                    response_message += content_message["text"]
+                elif "refusal" in content_message:
+                    response_message += content_message["refusal"]
+            ret["content"] = response_message
         ret["content"] += json.dumps(message.model_dump()["function_call"])
     return ret
 

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "instructor"
-version = "1.3.7"
+version = "1.4.0"
 description = "structured outputs for llm"
 authors = ["Jason Liu <jason@jxnl.co>"]
 license = "MIT"
@@ -10,7 +10,7 @@ repository = "https://github.com/jxnl/instructor"
 
 [tool.poetry.dependencies]
 python = "^3.9"
-openai = "^1.1.0"
+openai = "^1.40.0"
 pydantic = "^2.8.0"
 docstring-parser = "^0.16"
 typer = ">=0.9.0,<1.0.0"

tests/llm/test_openai/util.py

@@ -3,4 +3,5 @@
 models = ["gpt-4o-mini"]
 modes = [
     instructor.Mode.TOOLS,
+    instructor.Mode.STRUCTURED_OUTPUTS,
 ]