Spaces:
Runtime error
Runtime error
| """Regenerate terse evaluations and rebuild strategy for past reflections.""" | |
| import json | |
| from pathlib import Path | |
| from dotenv import load_dotenv | |
| import gemini | |
| load_dotenv() | |
| WORK_DIR = Path(__file__).parent | |
| RESULTS_PATH = WORK_DIR / "results.json" | |
| STRATEGY_PATH = WORK_DIR / "strategy.md" | |
| def _eval_prompt(entry: dict) -> str: | |
| home, away = entry["home"], entry["away"] | |
| match = entry["match"] | |
| prediction = entry.get("prediction", {}) | |
| result = entry["result"] | |
| pick = prediction.get("pick", "home") | |
| correct = pick == result["winner"] | |
| research_summary = "" | |
| for svc in entry.get("services_used", []): | |
| research_summary += f"\n- {svc['source']} (${svc['cost']:.4f}): {svc.get('reason', '')}" | |
| mistake_section = "" if correct else """MISTAKE: | |
| - One sentence: what the pick got wrong. | |
| - One sentence: the specific reasoning error. | |
| """ | |
| return f"""Write a terse post-match learning note for a prediction agent. | |
| Match: {match} | |
| Result: {home} {result['home_score']}-{result['away_score']} {away} (winner: {result['winner']}) | |
| Predicted: {pick} (confidence {prediction.get('confidence')}/10) — {'CORRECT' if correct else 'WRONG'} | |
| Reasoning given: {prediction.get('reasoning')} | |
| Research purchased (${entry.get('research_cost', 0):.4f}):{research_summary if research_summary else " none"} | |
| Rules: | |
| - Plain text only. No preamble, praise, or filler ("Here is...", "honest evaluation", etc.). | |
| - No markdown headers, bold, or horizontal rules. | |
| - Max 120 words. | |
| - Be blunt and specific. Name the exact mistake, not abstract narratives. | |
| Format exactly: | |
| {mistake_section}DATA: | |
| - useful: [source] — one line on what it told us that mattered | |
| - useless: [source] — one line on why it didn't help | |
| List only endpoints that were actually purchased. 2-5 bullets total across useful/useless.""" | |
| def _reflect_prompt(strategy: str, evaluation: str, match: str) -> str: | |
| return f"""Update the agent's strategy document. Write like engineering notes — terse, direct, no marketing language or dramatic rule names. | |
| Current strategy: | |
| {strategy if strategy else "(empty — first match)"} | |
| Latest match note ({match}): | |
| {evaluation} | |
| Rewrite with EXACTLY ONE rule change from this match. Keep still-valid existing rules. | |
| Format: | |
| # Strategy | |
| ## Rules | |
| - MATCH: one sentence rule | |
| ## Data sources | |
| - use: source type — why (MATCH) | |
| - skip: source type — why (MATCH) | |
| Max 250 words. Return ONLY the document.""" | |
| def main() -> None: | |
| results = json.loads(RESULTS_PATH.read_text()) | |
| reflected = sorted( | |
| [r for r in results if r.get("reflected")], | |
| key=lambda r: r.get("reflected_at") or "", | |
| ) | |
| if not reflected: | |
| print("No reflected entries to backfill.") | |
| return | |
| strategy = "" | |
| by_match = {r["match"]: r for r in results} | |
| for entry in reflected: | |
| match = entry["match"] | |
| print(f"Regenerating evaluation: {match}") | |
| evaluation = gemini.generate(_eval_prompt(entry)).strip() | |
| entry["evaluation"] = evaluation | |
| by_match[match]["evaluation"] = evaluation | |
| print(f" → {len(evaluation)} chars") | |
| print(f"Rebuilding strategy after: {match}") | |
| strategy = gemini.generate(_reflect_prompt(strategy, evaluation, match)).strip() | |
| entry["strategy_after"] = strategy | |
| by_match[match]["strategy_after"] = strategy | |
| STRATEGY_PATH.write_text(strategy + "\n") | |
| RESULTS_PATH.write_text(json.dumps(results, indent=2)) | |
| print(f"\nDone. Updated {len(reflected)} reflections and strategy.md ({len(strategy)} chars).") | |
| if __name__ == "__main__": | |
| main() | |