How the AI assistant guides recipe building — sees your schema and data, suggests stages and configuration, records every decision
The AI assistant incorporates a set of purpose-scoped chat interfaces, each tailored to a specific stage of the recipe-building workflow: data exploration, method selection, filter and transformation design, visualisation choice, and result interpretation.
Each interface has access to your actual schema and data — the real column names, types, and value distributions. Its suggestions are grounded in your specific data, not generic advice. The AI also adapts its conversational style to match an assessment of your skill level, drawn from the questions you ask and the answers you provide.
Every AI suggestion is recorded alongside your decisions. This provides a complete audit trail that purely AI-generated approaches lack. Every step of the recipe is explicit and inspectable. You can see what the AI suggested, what you chose, and why.
The AI-guided features can use cloud AI providers (Anthropic, OpenAI-compatible services like DeepSeek) or a local LLM via Ollama. With a local LLM, no data ever leaves your machine — preserving the local-first privacy architecture.
This is opt-in, not required. The structured recipe builder works without any AI connection. AI guidance is an enhancement, not a dependency.
The multi-turn AI recipe builder has been built and is functional — a conversational concierge that guides users through each stage of recipe construction, seeing their schema and data at every step. It is currently undergoing usability testing. The primary challenge is keeping the AI focused on the task at hand: large language models tend to wander, offering suggestions outside the current stage or drifting into general advice when they should be proposing a specific next step. This is an active area of refinement — the pipeline is correct, but the conversational steering needs tuning.