feat(instructor): introduce ANTHROPIC_JSON mode

.github/workflows/evals.yml

@@ -3,11 +3,11 @@ name: Weekly Tests
 on:
   workflow_dispatch:
   schedule:
-    - cron: '0 0 * * 0'  # Runs at 00:00 UTC every Sunday
+    - cron: "0 0 * * 0" # Runs at 00:00 UTC every Sunday
   push:
-    branches: [ main ]
+    branches: [main]
     paths-ignore:
-      - '**'  # Ignore all paths to ensure it only triggers on schedule
+      - "**" # Ignore all paths to ensure it only triggers on schedule
 
 jobs:
   weekly-tests:
@@ -20,15 +20,15 @@ jobs:
         uses: actions/setup-python@v4
         with:
           python-version: 3.11
-          cache: 'poetry'
+          cache: "poetry"
 
       - name: Install Poetry
         uses: snok/install-poetry@v1.3.1
 
       - name: Install dependencies
-        run: poetry install --with dev
+        run: poetry install --with dev,anthropic
 
       - name: Run all tests
         run: poetry run pytest tests/
         env:
-          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

.github/workflows/ruff.yml

@@ -3,7 +3,7 @@ name: Ruff
 on:
   push:
   pull_request:
-    branches: [ main ]
+    branches: [main]
 
 env:
   WORKING_DIRECTORY: "."
@@ -42,4 +42,4 @@ jobs:
         uses: actions/upload-artifact@v3
         with:
           name: ruff-log
-          path: ${{ env.WORKING_DIRECTORY }}/${{ env.RUFF_OUTPUT_FILENAME }}
+          path: ${{ env.WORKING_DIRECTORY }}/${{ env.RUFF_OUTPUT_FILENAME }}

.github/workflows/test.yml

@@ -33,21 +33,23 @@ jobs:
         uses: snok/install-poetry@v1.3.1
 
       - name: Install dependencies
-        run: poetry install --with dev anthropic
+        run: poetry install --with dev,anthropic
 
       - name: Run tests
-        run: poetry run pytest tests/ -k "not openai"
+        run: poetry run pytest tests/ -k "not llm"
         env:
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
 
       - name: Generate coverage report
         if: matrix.python-version == '3.11'
         run: |
-          poetry run coverage run -m pytest tests/ -k "not openai"
+          poetry run coverage run -m pytest tests/ -k "not llm"
           poetry run coverage report
           poetry run coverage html
         env:
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
 
       - name: Coveralls GitHub Action
         if: matrix.python-version == '3.11'

.github/workflows/test_docs.yml

@@ -10,7 +10,7 @@ jobs:
 
     strategy:
       matrix:
-        python-version: ['3.11']
+        python-version: ["3.11"]
 
     steps:
       - uses: actions/checkout@v2
@@ -22,8 +22,8 @@ jobs:
         uses: actions/setup-python@v4
         with:
           python-version: ${{ matrix.python-version }}
-          cache: 'poetry'
-      
+          cache: "poetry"
+
       - name: Cache Poetry virtualenv
         uses: actions/cache@v2
         with:
@@ -33,9 +33,9 @@ jobs:
             ${{ runner.os }}-poetry-
 
       - name: Install dependencies
-        run: poetry install --with dev,docs,test-docs
+        run: poetry install --with dev,docs,test-docs,anthropic
 
       - name: Run tests
-        run: poetry run pytest tests/openai/docs
+        run: poetry run pytest tests/llm/test_openai/docs
         env:
-          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

docs/blog/posts/anthropic.md

@@ -9,7 +9,7 @@ authors:
 
 A special shoutout to [Shreya](https://twitter.com/shreyaw_) for her contributions to the anthropic support. As of now, all features are operational with the exception of streaming support.
 
-For those eager to experiment, simply patch the client with `ANTHROPIC_TOOLS`, which will enable you to leverage the `anthropic` client for making requests.
+For those eager to experiment, simply patch the client with `ANTHROPIC_JSON`, which will enable you to leverage the `anthropic` client for making requests.
 
 ```
 pip install instructor[anthropic]
@@ -28,7 +28,7 @@ import instructor
 # Patching the Anthropics client with the instructor for enhanced capabilities
 anthropic_client = instructor.patch(
     create=anthropic.Anthropic().messages.create,
-    mode=instructor.Mode.ANTHROPIC_TOOLS
+    mode=instructor.Mode.ANTHROPIC_JSON
 )
 
 class Properties(BaseModel):

docs/concepts/patching.md

@@ -77,3 +77,33 @@ from openai import OpenAI
 
 client = instructor.patch(OpenAI(), mode=instructor.Mode.MD_JSON)
 ```
+
+## Anthropic JSON Mode
+
+Anthropic JSON mode uses Anthropic's JSON format for responses by setting the `mode` parameter to `instructor.Mode.ANTHROPIC_JSON` when patching the client.
+
+```python
+import instructor
+from anthropic import Anthropic
+
+create = instructor.patch(
+    create=anthropic.Anthropic().messages.create,
+    mode=instructor.Mode.ANTHROPIC_JSON
+)
+
+class User(BaseModel):
+    name: str
+    age: int
+
+resp = create(
+    model="claude-3-haiku-20240307",
+    max_tokens=1024,
+    messages=[
+        {
+            "role": "user",
+            "content": "Create a user",
+        }
+    ],
+    response_model=User,
+)
+```

examples/match_language/run_v1.py

@@ -1,4 +1,4 @@
-from pydantic import BaseModel, Field
+from pydantic import BaseModel
 from instructor import patch
 from openai import AsyncOpenAI
 from langdetect import detect

instructor/function_calls.py

@@ -99,7 +99,13 @@ def from_response(
             assert hasattr(completion, "content")
             return xml_to_model(cls, extract_xml(completion.content[0].text))  # type:ignore
 
-        assert hasattr(completion, "choices")
+        if mode == Mode.ANTHROPIC_JSON:
+            assert hasattr(completion, "content")
+            text = completion.content[0].text  # type: ignore
+            extra_text = extract_json_from_codeblock(text)
+            return cls.model_validate_json(extra_text)
+
+        assert hasattr(completion, "choices"), "No choices in completion"
 
         if completion.choices[0].finish_reason == "length":
             logger.error("Incomplete output detected, should increase max_tokens")

instructor/mode.py

@@ -13,6 +13,7 @@ class Mode(enum.Enum):
     MD_JSON = "markdown_json_mode"
     JSON_SCHEMA = "json_schema_mode"
     ANTHROPIC_TOOLS = "anthropic_tools"
+    ANTHROPIC_JSON = "anthropic_json"
 
     def __new__(cls, value: str) -> "Mode":
         member = object.__new__(cls)

instructor/process_response.py

@@ -11,7 +11,7 @@
 from openai.types.chat import ChatCompletion
 from pydantic import BaseModel
 
-
+import json
 import inspect
 import logging
 from typing import (
@@ -245,7 +245,7 @@ def handle_response_model(
                 As a genius expert, your task is to understand the content and provide
                 the parsed objects in json that match the following json_schema:\n
 
-                {response_model.model_json_schema()}
+                {json.dumps(response_model.model_json_schema(), indent=2)}
 
                 Make sure to return an instance of the JSON, not the schema itself
                 """
@@ -305,6 +305,34 @@ def handle_response_model(
                 new_kwargs["system"] = f"{system_prompt}\n{new_kwargs['system']}"
             else:
                 new_kwargs["system"] = system_prompt
+        elif mode == Mode.ANTHROPIC_JSON:
+            # anthropic wants system message to be a string so we first extract out any system message
+            openai_system_messages = [
+                message["content"]
+                for message in new_kwargs.get("messages", [])
+                if message["role"] == "system"
+            ]
+
+            new_kwargs["system"] = (
+                new_kwargs.get("system", "")
+                + "\n\n"
+                + "\n\n".join(openai_system_messages)
+            )
+
+            new_kwargs["system"] += f"""
+            You must only response in JSON format that adheres to the following schema:
+
+            <JSON_SCHEMA>
+            {json.dumps(response_model.model_json_schema(), indent=2)}
+            </JSON_SCHEMA>
+            """
+            new_kwargs["system"] = dedent(new_kwargs["system"])
+
+            new_kwargs["messages"] = [
+                message
+                for message in new_kwargs.get("messages", [])
+                if message["role"] != "system"
+            ]
         else:
             raise ValueError(f"Invalid patch mode: {mode}")