Use Logfire managed variables for prompt instructions

2026-04-ai-engineer/.claude/settings.json

@@ -0,0 +1,7 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(uv run:*)"
+    ]
+  }
+}

2026-04-ai-engineer/.gitignore

@@ -0,0 +1,3 @@
+mps/
+mps.tar.gz
+download.py

2026-04-ai-engineer/.python-version

@@ -0,0 +1 @@
+3.14

2026-04-ai-engineer/Makefile

@@ -0,0 +1,38 @@
+.DEFAULT_GOAL := main
+
+.PHONY: .uv
+.uv: ## Check that uv is installed
+	@uv --version || echo 'Please install uv: https://docs.astral.sh/uv/getting-started/installation/'
+
+.PHONY: install
+install: .uv ## Install python dependencies
+	# --only-dev to avoid building the python package, use make dev for that
+	uv sync
+
+.PHONY: format
+format: ## Format Python code
+	uv run ruff format
+	uv run ruff check --fix --fix-only
+
+.PHONY: lint
+lint: ## Lint Python code with ruff
+	uv run ruff format --check
+	uv run ruff check
+	uv run basedpyright
+
+.PHONY: main
+main: format lint ## run formatting and linting
+
+# (must stay last!)
+.PHONY: help
+help: ## Show this help (usage: make help)
+	@echo "Usage: make [recipe]"
+	@echo "Recipes:"
+	@awk '/^[a-zA-Z0-9_-]+:.*?##/ { \
+	    helpMessage = match($$0, /## (.*)/); \
+	        if (helpMessage) { \
+	            recipe = $$1; \
+	            sub(/:/, "", recipe); \
+	            printf "  \033[36mmake %-20s\033[0m %s\n", recipe, substr($$0, RSTART + 3, RLENGTH); \
+	    } \
+	}' $(MAKEFILE_LIST)

2026-04-ai-engineer/README.md

@@ -0,0 +1,53 @@
+# Prompt Optimization with GEPA and Cached Wikipedia Pages
+
+This example demonstrates automated prompt optimization using GEPA (Genetic-Pareto Prompt Evolution)
+with pydantic-ai and pydantic-evals, using cached Wikipedia pages for all 650 UK MPs.
+
+## Overview
+
+The example shows how to:
+
+- Generate a golden evaluation dataset from cached MP pages with a strong model
+- Load evaluation cases from JSON instead of hand-authoring them in code
+- Evaluate either all political relatives or just ancestor/parent-generation relatives
+- Build a GEPA adapter that integrates pydantic-evals with GEPA
+- Resolve agent instructions from a Logfire managed variable
+- Use a local Logfire variable provider so GEPA can optimize the same runtime-configurable value
+- Run automated prompt optimization that improves based on evaluation feedback
+
+## Running the Example
+
+```bash
+# Sync dependencies
+uv sync
+
+# Generate a golden dataset for the first 100 MPs
+uv run -m main generate-cases --limit 100 --model openai:gpt-5
+
+# Evaluate ancestor-only extraction on the test split
+uv run -m main eval --split test --focus ancestors --prompt-style initial
+
+# Compare initial vs expert prompts on the same task/model
+uv run -m main compare --split test --focus ancestors
+
+# Run optimization using train/val splits from the generated file
+uv run -m main optimize --train-split train --val-split val --focus ancestors --max-calls 50
+```
+
+## Files
+
+- `task.py` - Extraction schema, page preprocessing, relation filtering, and agent definition
+- `cases.py` - Golden dataset generation and JSON persistence
+- `evals.py` - Dataset loading and evaluation metrics
+- `adapter.py` - GEPA adapter that bridges pydantic-evals with GEPA
+- `main.py` - CLI script for running evaluation and optimization
+
+## Notes
+
+- The MP pages are read from the local `mps/` archive, not from live Wikipedia requests.
+- `generate-cases` is resumable. Re-run it with a higher `--limit`, `--all`, or a different `--output`.
+- The generated file stores the full set of political relatives. Evaluation can then filter to
+  `--focus ancestors` without regenerating the golden data.
+- The CLI configures Logfire with a local managed-variable provider. The agent reads its
+  instructions from `relations_instructions`, so the same code path can use a remote managed
+  variable in a deployed server and a local provider during GEPA optimization.