mirror of
https://github.com/escalante29/healthy-fit.git
synced 2026-03-21 15:08:46 +01:00
184c8330a7
Introduces DSPy-based nutrition and plan generation modules, including image analysis for nutritional info and personalized diet/exercise plans. Adds new API endpoints for health metrics/goals, nutrition image analysis, and plan management. Updates models, schemas, and backend structure to support these features, and includes initial training data and configuration for prompt optimization.
59 lines
2.2 KiB
Python
59 lines
2.2 KiB
Python
from dspy.teleprompt import BootstrapFewShotWithRandomSearch
|
|
|
|
from app.ai.nutrition import nutrition_module
|
|
from app.core.ai_config import configure_dspy
|
|
from scripts.nutrition_data import train_examples
|
|
|
|
# 0. Configure DSPy
|
|
configure_dspy()
|
|
|
|
|
|
# 1. Define Advanced Metric
|
|
def validate_nutrition_v2(example, pred, trace=None):
|
|
# Condition A: Accuracy (within 15% of ground truth)
|
|
actual_cals = example.nutritional_info.calories
|
|
pred_cals = pred.nutritional_info.calories
|
|
|
|
threshold = 0.15
|
|
lower = actual_cals * (1 - threshold)
|
|
upper = actual_cals * (1 + threshold)
|
|
is_accurate_count = lower <= pred_cals <= upper
|
|
|
|
# Condition B: Consistency (Macros match Calories within 20%)
|
|
# This prevents "hallucinated" numbers that don't satisfy physics
|
|
p = pred.nutritional_info.protein
|
|
c = pred.nutritional_info.carbs
|
|
f = pred.nutritional_info.fats
|
|
|
|
calculated_cals = (p * 4) + (c * 4) + (f * 9)
|
|
# Using a slightly looser bounds (20%) for fiber/rounding
|
|
consistency_threshold = 0.20
|
|
is_consistent_math = abs(calculated_cals - pred_cals) < (pred_cals * consistency_threshold)
|
|
|
|
# We want BOTH to be true
|
|
return is_accurate_count and is_consistent_math
|
|
|
|
|
|
# 2. Setup Advanced Optimizer
|
|
# RandomSearch is more expensive but finds better reasoning traces by randomizing
|
|
# the selection of few-shot examples.
|
|
# num_candidate_programs=10 means it will try 10 different combinations of prompts/examples
|
|
print("Configuring RandomSearch Optimizer...")
|
|
teleprompter = BootstrapFewShotWithRandomSearch(
|
|
metric=validate_nutrition_v2,
|
|
max_bootstrapped_demos=4,
|
|
max_labeled_demos=4,
|
|
num_candidate_programs=5, # Reduced to 5 for speed in this demo, typically 10-20
|
|
num_threads=1, # Sequential for stability, increase for parallelism
|
|
)
|
|
|
|
# 3. Compile (Optimize) the Module
|
|
print("Optimizing V2 (this includes random search and macro checks)...")
|
|
# Note: assertions are compiled into the pipeline automatically in newer DSPy,
|
|
# acting as soft constraints during the search.
|
|
compiled_nutrition = teleprompter.compile(nutrition_module, trainset=train_examples)
|
|
|
|
# 4. Save
|
|
compiled_nutrition.save("app/ai/nutrition_compiled.json")
|
|
print("Optimization V2 complete! Overwrote app/ai/nutrition_compiled.json")
|