from arcade_evals import ( BinaryCritic, EvalRubric, EvalSuite, ExpectedToolCall, SimilarityCritic, tool_eval, ) from arcade_tdk import ToolCatalog import arcade_google_docs from arcade_google_docs.enum import DocumentFormat, OrderBy from arcade_google_docs.tools import ( create_blank_document, create_document_from_text, get_document_by_id, insert_text_at_end_of_document, search_and_retrieve_documents, search_documents, ) # Evaluation rubric rubric = EvalRubric( fail_threshold=0.9, warn_threshold=0.95, ) catalog = ToolCatalog() catalog.add_module(arcade_google_docs) @tool_eval() def docs_eval_suite() -> EvalSuite: """Create an evaluation suite for Google Docs tools.""" suite = EvalSuite( name="Google Docs Tools Evaluation", system_message="You are an AI assistant that can create and manage Google Docs using the provided tools.", catalog=catalog, rubric=rubric, ) # A previous tool call to list_documents additional_messages = [ {"role": "user", "content": "list my 10 most recently created docs"}, { "role": "assistant", "content": "Please go to this URL and authorize the action: [Link](https://accounts.google.com/)", }, { "role": "assistant", "content": "", "tool_calls": [ { "id": "call_gegK723W2hXsORjBmq1Oexqk", "type": "function", "function": { "name": "Google_ListDocuments", "arguments": '{"limit":10,"order_by":"createdTime desc"}', }, } ], }, { "role": "tool", "content": '{"documents":[{"id":"1e0rCoT1Yd14WuuEvd3hSUcN_-VD3df4T3Q08uLm3TWc","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst10"},{"id":"1eTSWd-5zQds8K9OWYygwtCFMUyuuMize3bh3HaRsKts","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst9"},{"id":"19Dqugn0rVi89K0C__lpg1HbhQOTenccyZOhPgivTHMs","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst8"},{"id":"1RCibzx14eqP3vS9yI4nD13OKf8Vee56RiszS53OkR7I","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst7"},{"id":"1imFb04JQuBn8SiSsRFf6fEuYCyXkbII4KX8fsmnT0jo","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst6"},{"id":"1ZC3oypdfLWFgBd-emeSykJf9tZOae6USsFboygRCr-w","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst5"},{"id":"1-gFGNWmwLxEiKa6NNixLNq3X-phXRMORVZfVTfBg8Sc","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst4"},{"id":"1eQ8UBO_PY3Lem4R8OVdIc9ODXt0MrSUAnEu994Qz8P8","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst3"},{"id":"1TOxB0MLry-JzntDWDT1LFywTLdr3XDWPT5L5UsHMs5c","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst2"},{"id":"1a1UQ7C90s8kGfnO8k6wfAZz_Cy5nGN2MkCoRB5y2j3w","kind":"drive#file","mimeType":"application/vnd.google-apps.document","name":"Tst1"}],"documents_count":10}', "tool_call_id": "call_gegK723W2hXsORjBmq1Oexqk", "name": "Google_ListDocuments", }, { "role": "assistant", "content": "Here are your 10 most recently created Google Docs:\n\n1. [Tst10](https://docs.google.com/document/d/1e0rCoT1Yd14WuuEvd3hSUcN_-VD3df4T3Q08uLm3TWc)\n2. [Tst9](https://docs.google.com/document/d/1eTSWd-5zQds8K9OWYygwtCFMUyuuMize3bh3HaRsKts)\n3. [Tst8](https://docs.google.com/document/d/19Dqugn0rVi89K0C__lpg1HbhQOTenccyZOhPgivTHMs)\n4. [Tst7](https://docs.google.com/document/d/1RCibzx14eqP3vS9yI4nD13OKf8Vee56RiszS53OkR7I)\n5. [Tst6](https://docs.google.com/document/d/1imFb04JQuBn8SiSsRFf6fEuYCyXkbII4KX8fsmnT0jo)\n6. [Tst5](https://docs.google.com/document/d/1ZC3oypdfLWFgBd-emeSykJf9tZOae6USsFboygRCr-w)\n7. [Tst4](https://docs.google.com/document/d/1-gFGNWmwLxEiKa6NNixLNq3X-phXRMORVZfVTfBg8Sc)\n8. [Tst3](https://docs.google.com/document/d/1eQ8UBO_PY3Lem4R8OVdIc9ODXt0MrSUAnEu994Qz8P8)\n9. [Tst2](https://docs.google.com/document/d/1TOxB0MLry-JzntDWDT1LFywTLdr3XDWPT5L5UsHMs5c)\n10. [Tst1](https://docs.google.com/document/d/1a1UQ7C90s8kGfnO8k6wfAZz_Cy5nGN2MkCoRB5y2j3w)\n\nYou can click the links to open each document.", }, ] suite.add_case( name="Get document content", user_message="Can you read me the contents of Tst9 doc and also Tst10 doc please", expected_tool_calls=[ ExpectedToolCall( func=get_document_by_id, args={ "document_id": "1eTSWd-5zQds8K9OWYygwtCFMUyuuMize3bh3HaRsKts", }, ), ExpectedToolCall( func=get_document_by_id, args={ "document_id": "1e0rCoT1Yd14WuuEvd3hSUcN_-VD3df4T3Q08uLm3TWc", }, ), ], critics=[ BinaryCritic(critic_field="document_id", weight=0.6), ], additional_messages=additional_messages, ) suite.add_case( name="Insert text at end of document", user_message="Please add the text 'This is a new paragraph.' to the end of Tst4.", expected_tool_calls=[ ExpectedToolCall( func=insert_text_at_end_of_document, args={ "document_id": "1-gFGNWmwLxEiKa6NNixLNq3X-phXRMORVZfVTfBg8Sc", "text_content": "This is a new paragraph.", }, ) ], critics=[ BinaryCritic(critic_field="document_id", weight=0.5), SimilarityCritic(critic_field="text_content", weight=0.5), ], additional_messages=additional_messages, ) suite.add_case( name="Read the contents of two documents and then insert text at end of a different document.", user_message="Can you read me the contents of Tst9 doc and also Tst10 doc please. Also, please add the text 'This is a new paragraph.' to the end of Tst4.", expected_tool_calls=[ ExpectedToolCall( func=insert_text_at_end_of_document, args={ "document_id": "1-gFGNWmwLxEiKa6NNixLNq3X-phXRMORVZfVTfBg8Sc", "text_content": "This is a new paragraph.", }, ), ExpectedToolCall( func=get_document_by_id, args={ "document_id": "1eTSWd-5zQds8K9OWYygwtCFMUyuuMize3bh3HaRsKts", }, ), ExpectedToolCall( func=get_document_by_id, args={ "document_id": "1e0rCoT1Yd14WuuEvd3hSUcN_-VD3df4T3Q08uLm3TWc", }, ), ], critics=[ BinaryCritic(critic_field="document_id", weight=0.3), SimilarityCritic(critic_field="text_content", weight=0.3), ], additional_messages=additional_messages, ) suite.add_case( name="Create blank document", user_message="Create a new Doc titled 'Meeting Notes'.", expected_tool_calls=[ ExpectedToolCall( func=create_blank_document, args={ "title": "Meeting Notes", }, ) ], critics=[ SimilarityCritic(critic_field="title", weight=1.0), ], ) suite.add_case( name="Create document from text", user_message="Create a new doc called To-Do List with the content 'Buy groceries, Call mom, Finish report'.", expected_tool_calls=[ ExpectedToolCall( func=create_document_from_text, args={ "title": "To-Do List", "text_content": "Buy groceries\nCall mom\nFinish report", }, ) ], critics=[ SimilarityCritic(critic_field="title", weight=0.5), SimilarityCritic(critic_field="text_content", weight=0.5), ], ) suite.add_case( name="No tool call case", user_message="Create a new microsoft word document titled 'My Resume'.", expected_tool_calls=[], critics=[], ) return suite @tool_eval() def search_documents_eval_suite() -> EvalSuite: """Create an evaluation suite for Google Drive tools.""" suite = EvalSuite( name="Google Drive Tools Evaluation", system_message="You are an AI assistant that can manage Google Drive documents using the provided tools.", catalog=catalog, rubric=rubric, ) suite.add_case( name="Search documents in Google Drive", user_message="get my 49 most recently created documents, list the ones created most recently first.", expected_tool_calls=[ ExpectedToolCall( func=search_documents, args={ "order_by": [OrderBy.CREATED_TIME_DESC.value], "limit": 49, }, ) ], critics=[ BinaryCritic(critic_field="order_by", weight=0.5), BinaryCritic(critic_field="limit", weight=0.5), ], ) suite.add_case( name="Search documents in Google Drive based on document keywords", user_message="Search the documents that contain the word 'greedy' and the phrase 'hello, world'", expected_tool_calls=[ ExpectedToolCall( func=search_documents, args={ "document_contains": ["greedy", "hello, world"], }, ) ], critics=[ BinaryCritic(critic_field="document_contains", weight=1.0), ], ) suite.add_case( name="Search documents in a specific Google Drive based on document keywords", user_message="Search the documents that contain the word 'greedy' and the phrase 'hello, world' in the drive with id 'abc123'", expected_tool_calls=[ ExpectedToolCall( func=search_documents, args={ "document_contains": ["greedy", "hello, world"], "search_only_in_shared_drive_id": "abc123", }, ) ], critics=[ BinaryCritic(critic_field="search_only_in_shared_drive_id", weight=0.5), BinaryCritic(critic_field="document_contains", weight=0.5), ], ) suite.add_case( name="Search documents in a Google Drive Workspace organization domain based on document keywords", user_message="Search the documents that contain the phrase 'hello, world' in the organization domain", expected_tool_calls=[ ExpectedToolCall( func=search_documents, args={ "document_contains": ["hello, world"], "include_organization_domain_documents": True, }, ) ], critics=[ BinaryCritic(critic_field="include_organization_domain_documents", weight=0.5), BinaryCritic(critic_field="document_contains", weight=0.5), ], ) suite.add_case( name="Search documents in shared drives", user_message="Search the 5 documents from all drives corpora that nobody has touched in forever, excluding shared drives.", expected_tool_calls=[ ExpectedToolCall( func=search_documents, args={ "limit": 5, "include_shared_drives": False, }, ) ], critics=[ BinaryCritic(critic_field="include_shared_drives", weight=0.5), BinaryCritic(critic_field="limit", weight=0.5), ], ) suite.add_case( name="No tool call case", user_message="List my 10 most recently modified documents that are stored in my Microsoft OneDrive.", expected_tool_calls=[], critics=[], ) return suite @tool_eval() def search_and_retrieve_documents_eval_suite() -> EvalSuite: """Create an evaluation suite for Google Drive search and retrieve tools.""" suite = EvalSuite( name="Google Drive Tools Evaluation", system_message="You are an AI assistant that can manage Google Drive documents using the provided tools.", catalog=catalog, rubric=rubric, ) suite.add_case( name="Search and retrieve (write summary)", user_message="Write a summary of the documents in my Google Drive about 'MX Engineering'", expected_tool_calls=[ ExpectedToolCall( func=search_and_retrieve_documents, args={ "document_contains": ["MX Engineering"], "return_format": DocumentFormat.MARKDOWN, }, ) ], critics=[ BinaryCritic(critic_field="document_contains", weight=0.5), BinaryCritic(critic_field="return_format", weight=0.5), ], ) suite.add_case( name="Search and retrieve (project proposal)", user_message="Display the document contents in HTML format from my Google Drive that contain the phrase 'project proposal'.", expected_tool_calls=[ ExpectedToolCall( func=search_and_retrieve_documents, args={ "document_contains": ["project proposal"], "return_format": DocumentFormat.HTML, }, ) ], critics=[ BinaryCritic(critic_field="document_contains", weight=0.5), BinaryCritic(critic_field="return_format", weight=0.5), ], ) suite.add_case( name="Search and retrieve (meeting notes)", user_message="Retrieve documents that contain both 'meeting notes' and 'budget' in JSON format.", expected_tool_calls=[ ExpectedToolCall( func=search_and_retrieve_documents, args={ "document_contains": ["meeting notes", "budget"], "return_format": DocumentFormat.GOOGLE_API_JSON, }, ) ], critics=[ BinaryCritic(critic_field="document_contains", weight=0.5), BinaryCritic(critic_field="return_format", weight=0.5), ], ) suite.add_case( name="Search and retrieve (Q1 report)", user_message="Show me the content of the documents that mention 'Q1 report' but do not include the expression 'Project XYZ'.", expected_tool_calls=[ ExpectedToolCall( func=search_and_retrieve_documents, args={ "document_contains": ["Q1 report"], "document_not_contains": ["Project XYZ"], "return_format": DocumentFormat.MARKDOWN, }, ) ], critics=[ BinaryCritic(critic_field="document_contains", weight=1 / 3), BinaryCritic(critic_field="document_not_contains", weight=1 / 3), BinaryCritic(critic_field="return_format", weight=1 / 3), ], ) return suite