A framework for rapidly building large-scale, deterministic, interactive workflows with a fault-tolerant, conversational UX and AI-powered recommendations.

• Built on the principle on "Convention over configuration", ALA Ruby on Rails
• Uses:
  • A custom-built intent detection pipeline for fault-tolerant, self-correcting command routing
  • Pydantic and DSPy for parameter extraction and response generation

• Workflows are defined as a directory hierarchy of workitem types
  • Workitems can be ordered
  • Min/max constraints can be defined for the number of child workitems (one, unlimited, min/max)
  • Workflows can delegate to other workflows
• Commands are exposed for each workitem type
  • Commands may be specific to one workitem type or inheritable by child workitem types (base commands)
• Users are guided through the workflow but have complete control over navigation
  • Workflow navigation and command execution are exposed via a chat interface
  • Special constrained workflows are used to handle routing and parameter extraction errors
• AI-powered recommendations after every command interaction
  • Recommendations are generated AFTER a command has been processed. The user has complete control over the workflow and discretion over whether to follow a recommendation or take a different action.

• Clone the repo
  • Use WSL if you are on Windows
• Create a .env file in the passwords folder and add below keys if required
  • LITELLM_API_KEY_SYNDATA_GEN
  • LITELLM_API_KEY_PARAM_EXTRACTION
  • LITELLM_API_KEY_RESPONSE_GEN
  • LITELLM_API_KEY_AGENT
• Train fastworkflow, then train the sample workflow, finally run the sample workflow agent or assistant
  • Hint: review the .vscode/launch.json file for training/running the sample workflow

• AI enabled python applications
• Tools to enable rapid application development - declarative/imperative/visual