Constrain XLSX inference to structured table output

Use llama.cpp JSON-schema constrained generation for columns and rows, then construct the Gradio dataframe directly. This fixes malformed free-form CSV output on ZeroGPU.

app.py

@@ -1,4 +1,4 @@
-from io import StringIO
+import json
 from pathlib import Path
 import tomllib
 
@@ -190,8 +190,9 @@ def generate_xlsx_table(
         {
             "role": "system",
             "content": (
-                f"{system_prompt}\n\nReturn only valid CSV with one header row. "
-                "Do not wrap the CSV in a code fence or add prose."
+                f"{system_prompt}\n\nReturn only a JSON table object with a columns "
+                "array and a rows array. Every row must contain one string value per "
+                "column."
             ),
         },
         {
@@ -205,13 +206,34 @@ def generate_xlsx_table(
     ]
     completion = model.create_chat_completion(
         messages=messages,
+        response_format={
+            "type": "json_object",
+            "schema": {
+                "type": "object",
+                "properties": {
+                    "columns": {
+                        "type": "array",
+                        "items": {"type": "string"},
+                    },
+                    "rows": {
+                        "type": "array",
+                        "items": {
+                            "type": "array",
+                            "items": {"type": "string"},
+                        },
+                    },
+                },
+                "required": ["columns", "rows"],
+                "additionalProperties": False,
+            },
+        },
         max_tokens=1024,
         temperature=0.6,
         top_p=0.95,
         top_k=20,
     )
-    response = completion["choices"][0]["message"]["content"].rsplit("</think>", 1)[-1].strip()
-    return pd.read_csv(StringIO(response))
+    response = json.loads(completion["choices"][0]["message"]["content"])
+    return pd.DataFrame(response["rows"], columns=response["columns"])
 
 
 CSS = """