Add inference fallback for embedding errors

app.py

@@ -453,7 +453,8 @@ def _overall_html(result: Dict[str, Any]) -> str:
     for name in ["Negative", "Neutral", "Positive"]:
         val = _safe_float(probs.get(name))
         bars.append(f'<div class="prob-row"><span>{name}</span><div class="bar"><i style="width:{val * 100:.1f}%"></i></div><b>{val:.2f}</b></div>')
-    return f'<div class="overall-card"><div class="small-label">Overall sentiment</div><div class="overall-main">{_chip(label)} <span class="conf">confidence {conf:.2f}</span></div>{"".join(bars)}</div>'
+    fallback = '<div class="small-label">Fallback inference mode used for this sample.</div>' if result.get("fallback") else ""
+    return f'<div class="overall-card"><div class="small-label">Overall sentiment</div><div class="overall-main">{_chip(label)} <span class="conf">confidence {conf:.2f}</span></div>{"".join(bars)}{fallback}</div>'
 
 
 def _join_aspects(items: List[str]) -> str:

src/inference.py

@@ -71,6 +71,8 @@ _SENTIMENT_EVIDENCE_KEYWORDS = {
     "bad", "poor", "cheap", "terrible", "awful", "small", "large", "tight", "loose", "thin",
     "worth", "disappointed", "return", "returned", "recommend", "flattering", "beautiful",
 }
+_POSITIVE_HINTS = {"good", "great", "love", "loved", "perfect", "nice", "excellent", "comfortable", "soft", "worth", "recommend", "flattering", "beautiful"}
+_NEGATIVE_HINTS = {"bad", "poor", "cheap", "terrible", "awful", "small", "large", "tight", "loose", "thin", "disappointed", "return", "returned", "broke", "torn"}
 _STOPWORDS = {
     "a", "an", "the", "and", "or", "but", "if", "then", "than", "so", "as", "at", "by", "for", "from",
     "in", "into", "of", "on", "to", "with", "without", "is", "are", "was", "were", "be", "been", "being",
@@ -159,6 +161,65 @@ def _format_attention(out: Dict) -> Dict:
     return result
 
 
+def _fallback_prediction(review_text: str, product_meta: Dict, reason: str = "") -> Dict:
+    evidence = _text_evidence_by_aspect(review_text)
+    aspect_details = {}
+    aspect_labels = {}
+    pos_total = 0
+    neg_total = 0
+    for aspect in cfg.ASPECTS:
+        terms = evidence.get(aspect, [])
+        pos_hits = [t for t in terms if t in _POSITIVE_HINTS]
+        neg_hits = [t for t in terms if t in _NEGATIVE_HINTS]
+        if neg_hits and len(neg_hits) >= len(pos_hits):
+            label = "Negative"
+            confidence = 0.62
+            neg_total += 1
+        elif pos_hits:
+            label = "Positive"
+            confidence = 0.62
+            pos_total += 1
+        elif terms:
+            label = "Not_Mentioned"
+            confidence = 0.55
+        else:
+            label = "Not_Mentioned"
+            confidence = 0.60
+        aspect_labels[aspect] = label
+        aspect_details[aspect] = {
+            "label": label,
+            "confidence": confidence,
+            "class_probs": {
+                "Not_Mentioned": 0.70 if label == "Not_Mentioned" else 0.20,
+                "Positive": confidence if label == "Positive" else 0.20,
+                "Negative": confidence if label == "Negative" else 0.20,
+            },
+        }
+    overall_label = "Negative" if neg_total > pos_total else "Positive" if pos_total > 0 else "Neutral"
+    return {
+        "aspects": aspect_labels,
+        "aspect_details": aspect_details,
+        "overall": {
+            "label": overall_label,
+            "confidence": 0.60,
+            "class_probs": {"Negative": 0.60 if overall_label == "Negative" else 0.20,
+                            "Neutral": 0.60 if overall_label == "Neutral" else 0.20,
+                            "Positive": 0.60 if overall_label == "Positive" else 0.20},
+        },
+        "metadata_summary": _meta_summary(product_meta),
+        "meta_attention_by_aspect": {
+            aspect: {"features": 0.50, "categories": 0.25, "numeric": 0.25}
+            for aspect in cfg.ASPECTS
+        },
+        "top_meta_source_by_aspect": {
+            aspect: {"source": "features", "weight": 0.50}
+            for aspect in cfg.ASPECTS
+        },
+        "fallback": True,
+        "fallback_reason": reason or "Model embedding index guard activated.",
+    }
+
+
 def _slice_model_out(out: Dict, start: int, end: int) -> Dict:
     sliced = {}
     for k, v in out.items():
@@ -224,8 +285,11 @@ class AspectPredictor:
         meta_df = _meta_dict_to_df(product_meta)
         meta_vec = torch.from_numpy(self.meta_encoder.transform(meta_df)).float().to(self.device)
 
-        with torch.no_grad():
-            out = self.model(input_ids, attn_mask, meta_vec)
+        try:
+            with torch.no_grad():
+                out = self.model(input_ids, attn_mask, meta_vec)
+        except IndexError as exc:
+            return _fallback_prediction(review_text, product_meta, str(exc))
 
         preds = out["logits"][0].argmax(dim=-1).cpu().numpy()
         result = {
@@ -271,8 +335,17 @@ class AspectPredictor:
             )
             meta_vec = torch.from_numpy(self.meta_encoder.transform(meta_df)).float().to(self.device)
 
-            with torch.no_grad():
-                out = self.model(input_ids, attn_mask, meta_vec)
+            try:
+                with torch.no_grad():
+                    out = self.model(input_ids, attn_mask, meta_vec)
+            except IndexError as exc:
+                for item in chunk:
+                    results.append(_fallback_prediction(
+                        item.get("review_text", ""),
+                        item.get("product_meta", {}),
+                        str(exc),
+                    ))
+                continue
             preds = out["logits"].argmax(dim=-1).cpu().numpy()
 
             for j in range(len(chunk)):