Gemini 3 Integration

model = "gemini-3-flash-preview"

return genai.GenerationConfig(
    temperature=self.config.temperature,  # Always 1.0
)

EXPERIMENT_DESIGNER_SYSTEM_PROMPT = """You are an expert ML researcher designing experiments. Your goal is to systematically improve model performance through hypothesis-driven experimentation.

PRINCIPLES:
1. Each experiment tests a specific hypothesis derived from previous observations
2. Learn from both successes and failures in previous iterations
3. Consider data characteristics when selecting models and preprocessing
4. Apply appropriate preprocessing based on the data profile
5. Balance exploration (trying new approaches) with exploitation (refining what works)
6. Avoid repeating experiments that have already been tried
...
"""

{
  "experiment_name": "xgboost_tuned_learning_rate",
  "hypothesis": "Lower learning rate with more estimators may reduce overfitting",
  "model_type": "XGBRegressor",
  "model_params": {"n_estimators": 200, "learning_rate": 0.05, "max_depth": 6},
  "preprocessing": {
    "missing_values": "median",
    "scaling": "standard",
    "target_transform": "log"
  },
  "reasoning": "Previous iteration showed XGBoost overfit with learning_rate=0.1. Reducing to 0.05 and increasing n_estimators should improve generalization."
}

RESULTS_ANALYZER_SYSTEM_PROMPT = """You are an expert ML results analyst. Your role is to analyze experiment outcomes and provide actionable insights.

PRINCIPLES:
1. Compare current results against baseline, best, and previous experiments
2. Identify meaningful improvements versus noise
3. Detect patterns in model performance across iterations
4. Consider both the primary metric and secondary metrics
5. Note any anomalies or unexpected results
6. Provide observations that inform next experiment design

METRIC INTERPRETATION:
- Lower is better: RMSE, MSE, MAE, log_loss, error
- Higher is better: accuracy, f1, r2, precision, recall, AUC
...
"""

{
  "key_observations": [
    "XGBoost with lower learning rate (0.05) outperformed iteration 2's 0.1 by 10.3%",
    "Log transformation of target variable remains critical — iteration 1 without it had RMSE 2x worse",
    "Tree-based models (iterations 2-3) consistently outperform linear models (iteration 0)",
    "Diminishing returns observed — improvement from iteration 2 to 3 is only 2.1%"
  ],
  "reasoning": "The lower learning rate prevented overfitting while increased n_estimators maintained model capacity. However, we may be approaching the limit of this dataset's predictability."
}

HYPOTHESIS_GENERATOR_SYSTEM_PROMPT = """You are an expert ML researcher generating testable hypotheses. Your role is to analyze experiment results and propose the most promising directions for the next iteration.

PRINCIPLES:
1. Generate 2-3 ranked hypotheses based on the analysis
2. Each hypothesis should be specific and testable in a single experiment
3. Consider both exploration (trying new approaches) and exploitation (refining what works)
4. Reference specific metric values and patterns from the analysis
5. Suggest concrete model choices and hyperparameters when possible
6. Balance risk — include at least one "safe" refinement and one "exploratory" option

EXPLORATION VS EXPLOITATION:
- "explore": Early iterations or after plateau — try fundamentally different approaches
- "exploit": When a promising direction is found — refine parameters and preprocessing
- "balanced": Default — mix of both strategies
...
"""

if trend == "plateau" or iteration > max_iterations * 0.7:
    strategy_hint = "Consider more exploratory hypotheses — we may be in a local optimum."
elif trend == "improving" and iteration <= 5:
    strategy_hint = "The current direction is promising — consider both refining it and trying alternatives."
else:
    strategy_hint = "Balance exploration and exploitation."

{
  "analysis_summary": "XGBoost shows strong performance but diminishing returns suggest we're near optimal for this approach",
  "exploration_vs_exploitation": "balanced",
  "hypotheses": [
    {
      "hypothesis_id": "h1",
      "statement": "Fine-tune XGBoost regularization (alpha/lambda) to squeeze out remaining 1-2% improvement",
      "rationale": "Current model shows slight overfitting based on train/test gap",
      "suggested_model": "XGBRegressor",
      "suggested_params": {"reg_alpha": 0.5, "reg_lambda": 1.0},
      "confidence_score": 0.72,
      "priority": 1
    },
    {
      "hypothesis_id": "h2",
      "statement": "Try LightGBM as alternative gradient booster",
      "rationale": "LightGBM often matches XGBoost performance with faster training",
      "suggested_model": "LGBMRegressor",
      "confidence_score": 0.65,
      "priority": 2
    }
  ],
  "reasoning": "Focus on exploitation (h1) since we have a strong baseline, but keep one exploratory option (h2)."
}

REPORT_GENERATOR_SYSTEM_PROMPT = """You are an expert ML research writer creating a professional experiment report. Your role is to synthesize experiment results into a clear, insightful narrative.

PRINCIPLES:
1. Write in a professional, technical tone suitable for data science audiences
2. Be specific — reference actual metric values, model names, and iteration numbers
3. Explain WHY certain approaches worked or failed, not just WHAT happened
4. Connect insights across experiments to tell a coherent story
5. Provide actionable recommendations based on the evidence
...
"""

for attempt in range(self.config.max_retries):  # Default: 3 retries
    try:
        response = model.generate_content(content, generation_config=generation_config)
        return GeminiResponse(text=response.text, ...)
    
    except google_exceptions.ResourceExhausted as e:  # Rate limit (429)
        wait_time = self.config.retry_delay * (2**attempt)  # Exponential backoff
        time.sleep(wait_time)
    
    except google_exceptions.InvalidArgument as e:  # Bad request (400)
        raise GeminiError(f"Invalid request: {e}")  # Don't retry
    
    except Exception as e:  # Other errors
        wait_time = self.config.retry_delay * (2**attempt)
        time.sleep(wait_time)

def generate_json(self, prompt: str, ...) -> dict:
    # Add JSON instruction to prompt
    json_prompt = f"{prompt}\n\nRespond with valid JSON only. No additional text."
    
    response = self.generate(json_prompt, ...)
    
    # Clean up the response - remove markdown code blocks if present
    text = response.text.strip()
    if text.startswith("```json"):
        text = text[7:]
    if text.startswith("```"):
        text = text[3:]
    if text.endswith("```"):
        text = text[:-3]
    text = text.strip()
    
    return json.loads(text)