Tutorial: Spindle (AIP Logic & agents) editor¶
CSA Loom
aip-logiceditor — Spindle Studio, the Azure-native equivalent of Palantir AIP Logic and AIP agents: author typed AI logic and tool-calling agents over a Weave ontology, grounded on real data and run on Azure OpenAI / Azure AI Foundry. No Microsoft Fabric required.
What it is¶
AIP Logic builds no-code, typed LLM functions; AIP runs agents over the ontology. Spindle Studio covers both. You define a typed input schema and an ordered set of steps with dropdowns (no freeform JSON), bind a Weave ontology so the function grounds on its entity types and Lakehouse / Warehouse bindings, then run it two ways:
- Logic mode — a single grounded turn that writes real read-only T-SQL / Spark-SQL against Synapse and cites real rows, or
- Agent mode — a multi-step, tool-calling agent on the production copilot orchestrator with the full Loom data-tool registry.
You can also publish the logic as a real Azure AI Foundry Agent Service agent and inspect its per-step run trace.
When to use it¶
- You want a reusable, typed AI function (typed input → steps → typed output) grounded on governed data instead of an ad-hoc prompt.
- You want an agent that can call Loom data tools to answer multi-step questions over your ontology.
- You want to deploy the logic as a managed Foundry agent and audit its steps.
Step-by-step in Loom¶
- Create the item. Choose + New item → Spindle (AIP Logic & agents) (Fabric IQ). The editor opens at
/items/aip-logic/<id>. - Define typed inputs. Add named input parameters with types (string / number / boolean) in the field builder.
- Ground on the Weave. Bind a Weave ontology so Spindle runs against its entity types and Lakehouse / Warehouse bindings (real Synapse queries).
- Add ordered steps. Use Add step to add LLM-prompt, extract, or branch steps from a dropdown — no freeform JSON.
- Define the output. Set the typed output shape the function returns.
- Invoke. Flip the Logic / Agent switch, then run: Invoke function (single grounded turn) or Run agent (multi-step tool-calling). Both hit the live Azure OpenAI deployment; the agent returns a per-step run trace, or an honest remediation gate if no model is deployed.
- Publish as a Foundry agent. Use Publish to deploy the logic to Azure AI Foundry Agent Service, then run and inspect its steps — or use the Azure-native Invoke path where Agent Service is unsupported (for example Azure Government).
The Azure backend it rides on¶
- LLM: an Azure OpenAI deployment (
LOOM_AOAI_*); an unset deployment produces an honest gate naming the env var. - Grounding data: read-only Synapse SQL / Spark-SQL against the ontology's Lakehouse / Warehouse.
- Managed agents (optional): Azure AI Foundry Agent Service.
No Fabric required¶
Spindle grounds on a Loom Ontology and runs on Azure OpenAI / Foundry by default. Fabric Reflex / Fabric AI are never on the default path; honest gates name the exact AOAI or Foundry env var when a backend is unconfigured.
Learn more¶
- Ontology editor tutorial:
editor-ontology.md - Azure OpenAI: https://learn.microsoft.com/azure/ai-services/openai/overview
- Azure AI Foundry Agent Service: https://learn.microsoft.com/azure/ai-services/agents/overview