Create a New Task

Create a new Task with specified prompts, model configuration, and output format. Tasks are the core building blocks that define how the LLM processes inputs and generates structured outputs. Each Task includes: - System and user prompts that guide the LLM's behavior - Input parameters that can be dynamically provided when running the Task - Output format specification that ensures consistent, structured responses - Output modality supporting multiple formats: - `json`: Structured JSON responses (default) - `image`: AI-generated images (requires image-capable model) - `audio`: Voice/audio responses - `pdf`: Formatted PDF documents - `text`: Plain text output - `csv`: Structured CSV data tables - Optional fallback model for improved reliability when primary model fails - Optional task run visibility controls for managing access to task execution history - Optional image processing capabilities - Optional RAG (Retrieval Augmented Generation) for domain-specific knowledge **LLM Configuration (`llm_config`):** Model-specific configuration options. Common parameters include: - `temperature` (0.0-2.0): Controls randomness. Lower = more focused, higher = more creative - `top_p` (0.0-1.0): Nucleus sampling. Use either temperature OR top_p, not both - `max_tokens`: Maximum tokens in the response - `quality` (for image models): "standard" or "hd" - `size` (for image models): e.g., "1024x1024", "1792x1024" **Output Format (`output_format`):** Define the structure of task outputs. Two syntaxes are supported: *Shorthand syntax* (simple fields): ```json {"field_name": "str", "count": "int", "items": "array"} ``` *Object syntax* (with descriptions and constraints): ```json { "field_name": { "type": "str", "description": "What this field contains", "options": ["option1", "option2"] } } ``` Supported types: `str`/`string`, `int`/`integer`, `number`, `bool`/`boolean`, `array`, `object` **Task Run Visibility:** - `owner_only` (default): Only task owners can view all task runs - `editors_and_owners`: Task editors and owners can view all task runs - `all_viewers`: All users with view permission can see all task runs **File Input Mode (`file_input_mode`):** Specifies what file types are accepted for visual LLM processing: - `none` (default): No file input accepted - `image`: Image file required (PNG, JPEG, WebP, GIF) - `pdf`: PDF document required (requires PDF vision model) - `image_or_pdf`: Either image or PDF accepted - `any`: Any visually processable file accepted **PDF Vision Model Compatibility:** - Gemini 1.5/2.x/3.x models: Up to 50MB, 1000 pages - Claude 3.x/4.x models: Up to 30MB, 100 pages visual - GPT-4o models: Up to 20MB - GPT-3.5 / Groq: No PDF vision support **Common Use Cases:** *Structured Data Extraction:* Create tasks with JSON output to extract structured data from unstructured text (e.g., invoice parsing, contact extraction). *Image Generation:* Set `output_modality: "image"` with DALL-E or similar models to generate images from text prompts. *Document Analysis with Vision:* Set `file_input_mode` to accept images, PDFs, or both for visual analysis by the LLM. Use PDF-capable models (Gemini, Claude, GPT-4o) for direct PDF processing without text extraction. *Knowledge Base Q&A (RAG):* Configure `rag` with collection IDs to ground responses in your organization's documents and knowledge. *Automated Reports:* Use `output_modality: "pdf"` to generate formatted reports that can be downloaded or emailed. The Task will be created with an initial revision that becomes automatically active.