better handling of images

app.py

@@ -27,6 +27,52 @@ def _image_to_data_url(image_path: str, max_side: int = 1120, quality: int = 85)
     return "data:image/jpeg;base64," + base64.b64encode(buf.getvalue()).decode()
 
 
+def _normalize_content_for_api(content):
+    """
+    Gradio multimodal history uses {'type': 'file', 'file': FileData}; the HF
+    chat API expects {'type': 'image_url', 'image_url': {'url': ...}}.
+    """
+    if isinstance(content, str):
+        return content
+    if not isinstance(content, list):
+        return content
+    out: list = []
+    for part in content:
+        if isinstance(part, str):
+            out.append({"type": "text", "text": part})
+            continue
+        if not isinstance(part, dict):
+            continue
+        ptype = part.get("type")
+        if ptype == "text":
+            out.append({"type": "text", "text": part.get("text", "")})
+        elif ptype == "file":
+            fd = part.get("file")
+            if isinstance(fd, dict) and fd.get("path"):
+                out.append(
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": _image_to_data_url(fd["path"])},
+                    }
+                )
+        elif ptype == "image_url":
+            out.append(part)
+    if not out:
+        return ""
+    if len(out) == 1 and out[0].get("type") == "text":
+        return out[0].get("text", "")
+    return out
+
+
+def _normalize_history_message(msg: dict) -> dict:
+    role = msg.get("role")
+    if role not in ("user", "assistant", "system"):
+        return msg
+    content = msg.get("content")
+    normalized = _normalize_content_for_api(content)
+    return {**msg, "content": normalized}
+
+
 def respond(
     message,
     history: list[dict],
@@ -57,7 +103,7 @@ def respond(
         content = text
 
     messages = [{"role": "system", "content": system_message}]
-    messages.extend(history)
+    messages.extend(_normalize_history_message(m) for m in history)
     messages.append({"role": "user", "content": content})
 
     answer = run_agent(