How AI-guided recipe building works — concierge mode, suggestion engine, audit trail
The AI recipe builder follows a structured concierge pattern — not an open-ended chat. Each step presents a specific decision: choose input datasets, select a preparation operation, configure stage parameters, review the output, and proceed to the next stage. The AI observes your schema and data at every step and makes context-aware suggestions.
At each stage, the AI analyses the current dataset state and suggests: which stage type to add next, how to configure the chosen stage, and which columns to focus on. Suggestions appear as prompts you can accept, modify, or ignore. Each accepted suggestion is recorded in the audit trail alongside your decision.
Every AI suggestion and every user decision is logged with timestamps. The audit trail includes: the AI's original suggestion, whether you accepted or modified it, the final configuration, and the stage output preview. This creates a complete record of how the recipe was built — valuable for compliance, reproducibility, and review.
The AI concierge uses the same provider configuration as the AI Agents and AI-guided analysis features. Choose a cloud provider (DeepSeek, OpenAI, Anthropic) for fast, high-quality suggestions, or a local LLM (via Ollama or LM Studio) for complete data privacy. See AI Providers for setup instructions.
The recipe restructure in v0.12.0 gives the AI concierge a cleaner interface. The master-detail layout separates stage configuration (left panel) from live preview (right panel). The AI suggestion panel appears contextually — suggesting next steps based on the current pipeline state, not just the input schema.