Documentation

AI assistant.

The Odara ETL AI assistant: chat panel, provider settings, SQL and Python generation, pipeline generation and fixing errors.

11. AI Assistant

11.1 Chat panel

The AI Assistant is an in-editor chat panel (bottom-left purple button). You can ask it to generate pipelines, transforms, or fix errors. Open with the AI Assistant button or keyboard shortcut.

AI Assistant chat panel with Compose/Add Node/Modify tabs
AI Assistant — three modes: Compose (plan a pipeline from a goal), Add Node, Modify existing node

11.1.1 Three modes

The panel is split into three tabs. Compose writes a full pipeline from a natural-language goal; Add Node attaches a new node to existing ones; Modify edits the code of a node already on the canvas.

Compose tab with pipeline goal input
Compose — describe the goal (e.g., "Decrypt /tmp/sales.csv.pgp, filter amount > 100, load into Postgres") and Generate plan
Add Node tab connecting to existing nodes
Add Node — pick source nodes to connect from, choose SQL or Python, describe the transform
Modify tab rewriting an existing node
Modify — select a node on the canvas and describe the change (e.g., "add email validation, remove rows with invalid emails")

11.2 Settings (providers)

Supported providers (pick one in the AI Settings dialog):

  • DeepSeek (recommended for cost / code) — default model deepseek-chat.
  • OpenAI — GPT models.
  • Anthropic (Claude) — high-quality reasoning, good for complex pipelines.
  • Google (Gemini) — multimodal support.
  • MiniMax.

Open the gear icon in the AI Assistant header → pick a provider tab → paste its API key → choose a model (or Refresh models) → Test connection → save. Each provider's key is encrypted and stored separately, so switching providers later never loses a key. The status dot in the panel header turns green and shows the active provider/model.

11.3 Code generation (Add Node / Modify)

Code is generated from the AI Assistant panel, not from inside a node editor. Two modes:

  • Add Node — select one or more upstream source nodes, describe the transform, pick SQL or Python, and the AI adds a fully-written transform node wired to them. It sees the input columns/types, so the generated code references your real column names.
  • Modify — select an existing transform node and describe a change (e.g., "add email validation"); the AI rewrites that node's code in place.

Always review the generated code before running — it gets you most of the way, you own the final check (input table name, type casts, dependencies).

11.4 Pipeline generation (plan → compose)

From a natural-language goal like "Load yesterday's sales CSV into Postgres and email me a summary", the AI:

  1. Plans the steps as a high-level sketch.
  2. Composes each step into a fully configured node, wiring them together.
  3. Places the pipeline on the canvas for your review.

11.5 Fix errors with AI

When a node fails during execution, a Fix with AI button appears next to the error. The AI reads the error message + the node's config and suggests a corrected version. Accept or reject inline.