Physical Optimizer¶
The Physical Optimizer is responsible for selecting the concrete execution strategy for each operator in the logical plan. Its primary job is Model Selection.
Model Selection Strategy¶
Nirvana supports a hierarchy of models (e.g., from small/fast models like gpt-4o-mini to large/smart models like gpt-4). The physical optimizer determines the cheapest model that achieves acceptable quality for each specific operator.
Algorithm: Cascading Refinement¶
The optimizer iteratively tests models on a small sample of data:
- Baseline: Run the operator with the "Weakest" (cheapest) model and the "Strongest" (most expensive, ground truth) model.
- Comparison: Compare the outputs.
- If the Weakest model matches the Strongest model (high similarity/accuracy), use the Weakest model.
- If not, try the next model in the hierarchy.
- Improvement Margin: The optimizer selects a stronger model only if it provides a significant quality gain (above
improve_margin) relative to the cost increase.
Consistency Check¶
For Map operations, output consistency is often measured using Embedding Similarity. - It creates embeddings for the outputs of different models. - If the cosine similarity is high, the outputs are considered semantically equivalent, even if the text differs slightly.
Execution¶
The PhysicalOptimizer.optimize() method returns a plan where each operator node is annotated with the specific model to use during execution.