Update app.py

app.py

@@ -15,9 +15,9 @@ import torch
 from transformers import CLIPModel, CLIPProcessor
 
 # ========= CONFIG (override in Space → Settings → Variables) =========
-TEXT_MODEL_REPO = os.environ.get("TEXT_MODEL_REPO", "<your-username>/text-ft-food-rag")
-CLIP_MODEL_REPO = os.environ.get("CLIP_MODEL_REPO", "<your-username>/clip-ft-food-rag")
-DATASET_REPO    = os.environ.get("DATASET_REPO",    "<your-username>/food-rag-index")
+TEXT_MODEL_REPO = os.environ.get("TEXT_MODEL_REPO", "mamathew/text-ft-food-rag")
+CLIP_MODEL_REPO = os.environ.get("CLIP_MODEL_REPO", "mamathew/clip-ft-food-rag")
+DATASET_REPO    = os.environ.get("DATASET_REPO",    "mamathew/food-rag-index")
 
 # Inference API chat model (Gemma IT by default).
 LLM_ID = os.environ.get("LLM_ID", "google/gemma-2-2b-it")
@@ -57,7 +57,20 @@ except Exception:
     client = None
 
 # ---------------------- utils & dataclasses ----------------------
+from PIL import Image
+
+def _resolve_path(rel_or_abs: str) -> str:
+    # If relative, make it under the dataset snapshot root
+    p = rel_or_abs if os.path.isabs(rel_or_abs) else os.path.join(DATA_DIR, rel_or_abs)
+    # Resolve symlinks to a canonical path (helps in HF cache)
+    return os.path.realpath(p)
 
+def _open_image_safe(path: str):
+    try:
+        return Image.open(path).convert("RGB")
+    except Exception:
+        return None
+        
 def normalize_fa(s: str) -> str:
     if not s: return s
     return (s.replace("ي","ی").replace("ك","ک").replace("\u200c"," ").strip())
@@ -257,24 +270,30 @@ def call_llm(prompt: str) -> str:
 
 # ---------------------- gallery helpers ----------------------
 
-def display_gallery_pairs(pairs: List[Pair]) -> List[Tuple[str, str]]:
+def display_gallery_pairs(pairs):
     items = []
     for p in pairs:
-        if not p.image_path: continue
-        local_path = os.path.join(DATA_DIR, p.image_path) if not os.path.isabs(p.image_path) else p.image_path
+        if not p.image_path:
+            continue
+        local_path = _resolve_path(p.image_path)
         if os.path.exists(local_path):
-            caption = f"#{p.rank} — {p.title or ''}\nscore={(p.hscore if p.hscore is not None else p.score):.3f}"
-            items.append((local_path, caption))
+            img = _open_image_safe(local_path)
+            if img is not None:
+                caption = f"#{p.rank} — {p.title or ''}\nscore={(p.hscore if p.hscore is not None else p.score):.3f}"
+                items.append((img, caption))   # PIL image instead of path
     return items
 
-def display_gallery_images(img_hits: List[ImgHit]) -> List[Tuple[str, str]]:
+def display_gallery_images(img_hits):
     items = []
     for h in img_hits:
-        if not h.image_path: continue
-        local_path = os.path.join(DATA_DIR, h.image_path) if not os.path.isabs(h.image_path) else h.image_path
+        if not h.image_path:
+            continue
+        local_path = _resolve_path(h.image_path)
         if os.path.exists(local_path):
-            caption = f"#{h.rank} — {h.title or ''}\nscore={h.score:.3f}"
-            items.append((local_path, caption))
+            img = _open_image_safe(local_path)
+            if img is not None:
+                caption = f"#{h.rank} — {h.title or ''}\nscore={h.score:.3f}"
+                items.append((img, caption))   # PIL image instead of path
     return items
 
 # ---------------------- main app logic ----------------------
@@ -298,13 +317,13 @@ def answer(question: str, image: Optional[Image.Image], topk: int, k_ctx: int, u
         img_hits = search_image_by_text(question, topk=min(8, topk))
         gallery = display_gallery_images(img_hits)
 
-    top_image_path = gallery[0][0] if gallery else None
+    top_image = gallery[0][0] if gallery else None
 
     # Table
     def fmt(x): return "—" if x is None else f"{x:.3f}"
     table = [[p.rank, p.title or "", fmt(p.tscore), fmt(p.iscore), fmt(p.hscore or p.score), p.doc_id] for p in top_pairs]
 
-    return gen, table, gallery, top_image_path
+    return gen, table, gallery, top_image
 
 # ---------------------- UI ----------------------
 
@@ -325,13 +344,18 @@ with gr.Blocks() as demo:
     out_text   = gr.Textbox(label="پاسخ (Answer)")
     out_table  = gr.Dataframe(headers=["Rank", "Title", "Text S", "Image S", "Hybrid S", "Doc ID"], label="Top-K retrieval")
     out_gallery = gr.Gallery(label="تصاویر مرتبط (Image matches)", columns=5, height=240)
-    out_img_top = gr.Image(label="بهترین تصویر")
+    out_img_top = gr.Image(label="Top image match")")
 
     btn.click(
         answer,
         inputs=[q, img, topk, kctx, use_img, alpha],
         outputs=[out_text, out_table, out_gallery, out_img_top]
     )
-
+ALLOWED = [
+    DATA_DIR,
+    os.path.join(DATA_DIR, "data"),
+    os.path.join(DATA_DIR, "data", "interim"),
+    os.path.join(DATA_DIR, "data", "interim", "images_cache"),
+]
 if __name__ == "__main__":
-    demo.launch(allowed_paths=[DATA_DIR])
+    demo.launch(allowed_paths=[os.path.realpath(p) for p in ALLOWED])