Tune optimizer acceptance policy and add optimization modes

Improve edit acceptance by supporting conservative/balanced/aggressive modes, continuing after rejected steps, prioritizing goal-level gains, and surfacing rejection reasons in iteration logs.

Made-with: Cursor

models.py

@@ -90,6 +90,7 @@ class OptimizerRequest(BaseModel):
     max_iterations: int = 2
     candidates_per_iteration: int = 2
     temperature: float = 0.25
+    optimization_mode: str = "balanced"
 
 
 class OptimizerResponse(BaseModel):
@@ -99,4 +100,5 @@ class OptimizerResponse(BaseModel):
     final_metrics: Dict[str, Any] = Field(default_factory=dict)
     iterations: List[Dict[str, Any]] = Field(default_factory=list)
     applied_changes: int = 0
+    optimization_mode: str = "balanced"
     error: str = ""

optimizer.py

@@ -358,18 +358,59 @@ def _llm_rewrite_sentence(
     return str(parsed["revised_sentence"]).strip()
 
 
-def _is_candidate_valid(prev_metrics: Dict[str, Any], next_metrics: Dict[str, Any]) -> bool:
-    if next_metrics["bert_low_count"] > prev_metrics["bert_low_count"]:
-        return False
-    if next_metrics["bm25_remove_count"] > prev_metrics["bm25_remove_count"]:
-        return False
-    if next_metrics["semantic_gap_count"] > prev_metrics["semantic_gap_count"]:
-        return False
+def _goal_improved(goal_type: str, prev_metrics: Dict[str, Any], next_metrics: Dict[str, Any]) -> bool:
+    if goal_type == "bert":
+        return next_metrics["bert_low_count"] < prev_metrics["bert_low_count"]
+    if goal_type == "bm25":
+        return next_metrics["bm25_remove_count"] < prev_metrics["bm25_remove_count"]
+    if goal_type == "semantic":
+        return next_metrics["semantic_gap_count"] < prev_metrics["semantic_gap_count"]
+    if goal_type == "ngram":
+        return next_metrics["ngram_signal_count"] < prev_metrics["ngram_signal_count"]
+    return next_metrics["score"] > prev_metrics["score"]
+
+
+def _is_candidate_valid(
+    prev_metrics: Dict[str, Any],
+    next_metrics: Dict[str, Any],
+    goal_type: str,
+    optimization_mode: str,
+) -> Tuple[bool, List[str], bool]:
+    mode = (optimization_mode or "balanced").lower()
+    if mode not in {"conservative", "balanced", "aggressive"}:
+        mode = "balanced"
+
+    cfg = {
+        "conservative": {"max_score_drop": 0.0, "max_title_drop": 0.02},
+        "balanced": {"max_score_drop": 1.0, "max_title_drop": 0.03},
+        "aggressive": {"max_score_drop": 2.0, "max_title_drop": 0.05},
+    }[mode]
+
+    reasons = []
+    score_drop = float(prev_metrics["score"]) - float(next_metrics["score"])
+    if score_drop > cfg["max_score_drop"]:
+        reasons.append(f"score_drop>{cfg['max_score_drop']}")
+
+    # Hard regressions in critical counters.
+    if next_metrics["bm25_remove_count"] > prev_metrics["bm25_remove_count"] + (1 if mode == "aggressive" else 0):
+        reasons.append("bm25_remove_regression")
+    if next_metrics["bert_low_count"] > prev_metrics["bert_low_count"] + (1 if mode == "aggressive" else 0):
+        reasons.append("bert_low_regression")
+    if next_metrics["semantic_gap_count"] > prev_metrics["semantic_gap_count"] + (1 if mode == "aggressive" else 0):
+        reasons.append("semantic_gap_regression")
+
     prev_title = prev_metrics.get("title_bert_score")
     next_title = next_metrics.get("title_bert_score")
-    if prev_title is not None and next_title is not None and next_title < (prev_title - 0.03):
-        return False
-    return True
+    if prev_title is not None and next_title is not None and next_title < (prev_title - cfg["max_title_drop"]):
+        reasons.append("title_bert_drop")
+
+    improved = _goal_improved(goal_type, prev_metrics, next_metrics)
+
+    # In conservative mode require explicit goal improvement.
+    if mode == "conservative" and not improved:
+        reasons.append("goal_not_improved")
+
+    return (len(reasons) == 0), reasons, improved
 
 
 def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
@@ -390,6 +431,7 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
     candidates_per_iteration = int(request_data.get("candidates_per_iteration", 2) or 2)
     candidates_per_iteration = max(1, min(5, candidates_per_iteration))
     temperature = float(request_data.get("temperature", 0.25) or 0.25)
+    optimization_mode = str(request_data.get("optimization_mode", "balanced") or "balanced")
 
     baseline_analysis = _build_analysis_snapshot(
         target_text, competitors, keywords, language, target_title, competitor_titles
@@ -447,7 +489,9 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
                 )
                 cand_semantic = _build_semantic_snapshot(candidate_text, competitors, language)
                 cand_metrics = _compute_metrics(cand_analysis, cand_semantic, keywords, language)
-                valid = _is_candidate_valid(current_metrics, cand_metrics)
+                valid, invalid_reasons, goal_improved = _is_candidate_valid(
+                    current_metrics, cand_metrics, goal["type"], optimization_mode
+                )
                 delta_score = round(cand_metrics["score"] - current_metrics["score"], 3)
                 candidates.append(
                     {
@@ -458,6 +502,8 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
                         "semantic": cand_semantic,
                         "metrics": cand_metrics,
                         "valid": valid,
+                        "goal_improved": goal_improved,
+                        "invalid_reasons": invalid_reasons,
                         "delta_score": delta_score,
                     }
                 )
@@ -467,6 +513,8 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
                         "candidate_index": ci + 1,
                         "error": str(e),
                         "valid": False,
+                        "goal_improved": False,
+                        "invalid_reasons": [str(e)],
                         "delta_score": -999.0,
                     }
                 )
@@ -484,6 +532,8 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
                         {
                             "candidate_index": c.get("candidate_index"),
                             "valid": c.get("valid", False),
+                            "goal_improved": c.get("goal_improved", False),
+                            "invalid_reasons": c.get("invalid_reasons", []),
                             "delta_score": c.get("delta_score"),
                             "error": c.get("error"),
                         }
@@ -491,10 +541,15 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
                     ],
                 }
             )
-            break
+            continue
 
-        best = sorted(valid_candidates, key=lambda c: c["metrics"]["score"], reverse=True)[0]
-        if best["metrics"]["score"] <= current_metrics["score"]:
+        best = sorted(
+            valid_candidates,
+            key=lambda c: (1 if c.get("goal_improved") else 0, c["metrics"]["score"]),
+            reverse=True,
+        )[0]
+        # Accept candidate if it improves goal OR improves total score.
+        if not best.get("goal_improved") and best["metrics"]["score"] <= current_metrics["score"]:
             logs.append(
                 {
                     "step": step + 1,
@@ -506,7 +561,7 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
                     "current_score": current_metrics["score"],
                 }
             )
-            break
+            continue
 
         prev_metrics = current_metrics
         current_text = best["text"]
@@ -535,4 +590,5 @@ def optimize_text(request_data: Dict[str, Any]) -> Dict[str, Any]:
         "final_metrics": current_metrics,
         "iterations": logs,
         "applied_changes": applied_changes,
+        "optimization_mode": optimization_mode,
     }

templates/index.html

@@ -302,6 +302,14 @@
                                 <label class="form-label small text-muted mb-1">Temp</label>
                                 <input type="number" id="optimizerTemp" class="form-control" min="0" max="1.2" step="0.05" value="0.25">
                             </div>
+                            <div class="col-md-3">
+                                <label class="form-label small text-muted mb-1">Режим оптимизации</label>
+                                <select id="optimizerMode" class="form-select">
+                                    <option value="conservative">Conservative</option>
+                                    <option value="balanced" selected>Balanced</option>
+                                    <option value="aggressive">Aggressive</option>
+                                </select>
+                            </div>
                         </div>
                         <div class="d-flex gap-2 mt-3">
                             <button class="btn btn-dark" onclick="runLlmOptimization()">Запустить оптимизацию</button>
@@ -536,7 +544,8 @@
                 optimizer_model: document.getElementById('optimizerModel').value,
                 optimizer_iterations: Number(document.getElementById('optimizerIterations').value || 2),
                 optimizer_candidates: Number(document.getElementById('optimizerCandidates').value || 2),
-                optimizer_temperature: Number(document.getElementById('optimizerTemp').value || 0.25)
+                optimizer_temperature: Number(document.getElementById('optimizerTemp').value || 0.25),
+                optimizer_mode: document.getElementById('optimizerMode').value
             },
             state: {
                 analysis_result: currentData,
@@ -586,6 +595,7 @@
         document.getElementById('optimizerIterations').value = 2;
         document.getElementById('optimizerCandidates').value = 2;
         document.getElementById('optimizerTemp').value = 0.25;
+        document.getElementById('optimizerMode').value = 'balanced';
 
         // Competitor text fields
         const competitorsList = document.getElementById('competitorsList');
@@ -638,6 +648,7 @@
         document.getElementById('optimizerIterations').value = inp.optimizer_iterations ?? 2;
         document.getElementById('optimizerCandidates').value = inp.optimizer_candidates ?? 2;
         document.getElementById('optimizerTemp').value = inp.optimizer_temperature ?? 0.25;
+        document.getElementById('optimizerMode').value = inp.optimizer_mode || 'balanced';
 
         // Title character counter refresh
         const titleLen = (inp.target_title || '').length;
@@ -817,6 +828,7 @@
         const iterRows = (data.iterations || []).map(it => {
             const before = it.metrics_before ? it.metrics_before.score : '-';
             const after = it.metrics_after ? it.metrics_after.score : '-';
+            const reason = it.reason || (it.candidates ? 'all candidates rejected by constraints' : '-');
             return `<tr>
                 <td>${it.step}</td>
                 <td>${it.status}</td>
@@ -824,6 +836,7 @@
                 <td>${before}</td>
                 <td>${after}</td>
                 <td>${it.delta_score ?? '-'}</td>
+                <td>${reason}</td>
             </tr>`;
         }).join('');
 
@@ -831,6 +844,7 @@
             <div class="stat-card">
                 <h6 class="card-title">Результат оптимизации</h6>
                 <div class="small mb-2">Применено правок: <strong>${data.applied_changes || 0}</strong></div>
+                <div class="small mb-2">Режим: <strong>${data.optimization_mode || 'balanced'}</strong></div>
                 <div class="table-responsive">
                     <table class="table table-sm table-bordered mb-0">
                         <thead class="table-light"><tr><th>Метрика</th><th>До</th><th>После</th></tr></thead>
@@ -842,8 +856,8 @@
                 <h6 class="card-title">Лог итераций</h6>
                 <div class="table-responsive">
                     <table class="table table-sm table-hover mb-0">
-                        <thead><tr><th>#</th><th>Статус</th><th>Цель</th><th>Score до</th><th>Score после</th><th>Δ</th></tr></thead>
-                        <tbody>${iterRows || '<tr><td colspan="6" class="text-muted text-center">Нет данных</td></tr>'}</tbody>
+                        <thead><tr><th>#</th><th>Статус</th><th>Цель</th><th>Score до</th><th>Score после</th><th>Δ</th><th>Причина/комментарий</th></tr></thead>
+                        <tbody>${iterRows || '<tr><td colspan="7" class="text-muted text-center">Нет данных</td></tr>'}</tbody>
                     </table>
                 </div>
             </div>`;
@@ -882,7 +896,8 @@
             model: (document.getElementById('optimizerModel').value || '').trim(),
             max_iterations: Number(document.getElementById('optimizerIterations').value || 2),
             candidates_per_iteration: Number(document.getElementById('optimizerCandidates').value || 2),
-            temperature: Number(document.getElementById('optimizerTemp').value || 0.25)
+            temperature: Number(document.getElementById('optimizerTemp').value || 0.25),
+            optimization_mode: document.getElementById('optimizerMode').value || 'balanced'
         };
 
         document.getElementById('loader').style.display = 'block';