Predictive skills

skills/rai-discovery/SKILL.md

@@ -386,7 +386,12 @@ Each suggestion includes a `reasoners` field — an ordered list specifying the
 }
 ```
 
-After discovery, use the appropriate formulation skill for the chosen reasoner type.
+After discovery, route to the appropriate skill based on the `reasoners` tag:
+
+- **prescriptive** → `rai-prescriptive-problem-formulation`
+- **graph** → `rai-graph-analysis`
+- **rules** → `rai-rules-authoring`
+- **predictive** → `rai-predictive-modeling`
 
 ---
 

skills/rai-graph-analysis/SKILL.md

@@ -37,6 +37,7 @@ description: Graph algorithm selection and execution on PyRel v1 models. Covers
 - Ontology design decisions (concept modeling, data mapping) — see `rai-ontology-design`
 - Optimization formulation (variables, constraints, objectives) — see `rai-prescriptive-problem-formulation`
 - Business rule authoring (validation, classification, alerting) — see `rai-rules-authoring`
+- GNN graph construction for predictive pipelines — see `rai-predictive-modeling`
 
 **Overview (process steps):**
 1. Study the existing model — understand base definitions, coding conventions, and what's already wired

skills/rai-predictive-management/SKILL.md

@@ -0,0 +1,194 @@
+---
+name: rai-predictive-management
+description: Register trained GNN models to Snowflake Model Registry and load previously trained models for inference. Use after training with rai-predictive-training, or when loading a saved model by registry key or run ID.
+---
+
+# Predictive Management
+<!-- v1-SENSITIVE -->
+
+## Summary
+
+**What:** Encodes the model lifecycle workflow — registering trained GNN models and loading them for later inference.
+
+**When to use:**
+- Saving a trained model to Snowflake Model Registry
+- Loading a previously registered model by name and version
+- Loading a model by run ID
+- Understanding the train-register-load multi-session workflow
+
+**When NOT to use:**
+- Defining concepts, building graphs, or configuring features — see `rai-predictive-modeling`
+- Training a model or generating predictions — see `rai-predictive-training`
+
+**Overview:**
+1. Register a trained model with `gnn.register_model()`
+2. Load a model by registry key or run ID with `GNN(...).load()`
+3. Generate predictions with the loaded model
+
+---
+
+## Quick Reference
+
+**Register:**
+```python
+gnn.register_model(
+    model_database="DB",
+    model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector",
+    version_name="v1",
+)
+```
+
+**Load by registry key:**
+```python
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph, pt=pt,
+    model_database="DB", model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector", version_name="v1",
+)
+gnn.load()
+```
+
+**Load by run ID:**
+```python
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph, pt=pt,
+    model_run_id="01c2d9a0-0711-c54d-000a-1dc707f7a1e6",
+)
+gnn.load()
+```
+
+**What to include vs. omit when loading:**
+
+| Include | Omit |
+|---------|------|
+| `database`, `schema` | `train`, `validation` |
+| `exp_database`, `exp_schema` | `task_type`, `eval_metric` |
+| `graph`, `pt` | hyperparameters (`device`, `n_epochs`, etc.) |
+| model identifier (registry key or run ID) | |
+
+---
+
+## Register a Model
+
+After `gnn.fit()` completes, register the model to the Snowflake Model Registry:
+
+```python
+gnn.register_model(
+    model_database="DB",
+    model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector",
+    version_name="v1",
+    comment="Initial training run",  # optional
+)
+```
+
+The combination of `(model_database, model_schema, model_name, version_name)` uniquely identifies a registered model.
+
+---
+
+## Load a Model
+
+### By Registry Key
+
+```python
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph,
+    pt=pt,
+    model_database="DB",
+    model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector",
+    version_name="v1",
+)
+gnn.load()
+
+User.predictions = gnn.predictions(domain=Test)
+```
+
+### By Run ID
+
+```python
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph,
+    pt=pt,
+    model_run_id="01c2d9a0-0711-c54d-000a-1dc707f7a1e6",
+)
+gnn.load()
+
+User.predictions = gnn.predictions(domain=Test)
+```
+
+After `gnn.load()`, use `gnn.predictions(domain=Test)` exactly as after `gnn.fit()` — the prediction workflow is the same (see `rai-predictive-training`).
+
+---
+
+## Train-Register-Load Workflow
+
+A typical multi-session workflow:
+
+### Session 1: Train and Register
+
+```python
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph, pt=pt,
+    train=Train, validation=Val,
+    task_type="binary_classification", eval_metric="roc_auc",
+    has_time_column=True,
+    device="cuda", n_epochs=5,
+)
+gnn.fit()
+gnn.register_model(
+    model_database="DB", model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector", version_name="v1",
+)
+```
+
+### Session 2: Load and Predict
+
+Rebuild the graph and PropertyTransformer with the same structure as training, then load:
+
+```python
+# Rebuild graph and pt (same as training session)
+gnn_graph = Graph(model, directed=True, weighted=False, aggregator="sum")
+# ... define edges ...
+pt = PropertyTransformer(...)
+
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph, pt=pt,
+    model_database="DB", model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector", version_name="v1",
+)
+gnn.load()
+User.predictions = gnn.predictions(domain=Test)
+```
+
+---
+
+## Common Pitfalls
+
+| Mistake | Cause | Fix |
+|---------|-------|-----|
+| Calling `register_model()` before `fit()` | Model must be trained first — `RuntimeError`, no weights to serialize | Always call `gnn.fit()` before `gnn.register_model()` |
+| Omitting `graph` or `pt` when loading | Loaded models still need the graph structure — `RuntimeError` | Provide the same `graph` and `pt` used during training |
+| Passing `train`, `validation`, or hyperparameters when loading | These are training-only parameters — ignored or unexpected behavior | Omit `train`, `validation`, `task_type`, `eval_metric`, and all hyperparameters |
+| Reusing the same `(name, version)` tuple | Registry keys must be unique — `RegistryError` | Use a new `version_name` for each registration |
+
+---
+
+## Examples
+
+| Pattern | Description | File |
+|---------|-------------|------|
+| Register and load | Complete train-register-load workflow across sessions | [examples/register_and_load.py](examples/register_and_load.py) |

skills/rai-predictive-management/examples/register_and_load.py

@@ -0,0 +1,52 @@
+"""
+GNN Model Management — Register and Load Workflow
+==================================================
+Demonstrates the train-register-load pattern across sessions.
+
+Session 1: Train a model and register it to Snowflake Model Registry.
+Session 2: Load the registered model and generate predictions.
+"""
+
+# ── Session 1: Train and Register ───────────────────────────────────────────
+# Assumes data model from `rai-predictive-modeling`:
+#   gnn_graph, pt, Train, Val, Test, User
+
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph, pt=pt,
+    train=Train, validation=Val,
+    task_type="binary_classification", eval_metric="roc_auc",
+    has_time_column=True,
+    device="cuda", n_epochs=5,
+)
+gnn.fit()
+
+gnn.register_model(
+    model_database="DB",
+    model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector",
+    version_name="v1",
+    comment="Initial training run",
+)
+
+
+# ── Session 2: Load and Predict ─────────────────────────────────────────────
+# REQUIRED: Rebuild the same graph and PT structure used during training
+gnn_graph = Graph(model, directed=True, weighted=False, aggregator="sum")
+Edge = gnn_graph.Edge
+# ... define edges (same as training session) ...
+pt = PropertyTransformer(...)  # same config as training session
+
+gnn = GNN(
+    database="DB", schema="SCHEMA",
+    exp_database="DB", exp_schema="EXPERIMENTS",
+    graph=gnn_graph, pt=pt,
+    model_database="DB",
+    model_schema="MODEL_REGISTRY",
+    model_name="fraud_detector",
+    version_name="v1",
+)
+gnn.load()
+
+User.predictions = gnn.predictions(domain=Test)