perf: stream SQL to UI, trim few-shots 7→4, GPU duration 30s

- handle_query is now a generator: tokens stream into the SQL panel
within seconds instead of a blank wait
- prompts.py: drop 3 redundant few-shot examples (simple patterns
covered by remaining 4) to cut bnb-4bit prefill time
- spaces.GPU(duration=30): shorter durations get ZeroGPU queue priority
- max_tokens 192 (observed outputs are ~140-170 chars)

app.py

@@ -12,11 +12,13 @@ import gradio as gr
 # spaces.GPU is only available on HF Spaces — use a no-op locally
 try:
     import spaces
-    _gpu_decorator = spaces.GPU(duration=120)
+    # Short duration = higher priority in the ZeroGPU queue. Generation
+    # takes ~5s; 30s leaves ample headroom.
+    _gpu_decorator = spaces.GPU(duration=30)
 except ImportError:
     _gpu_decorator = lambda fn: fn  # no-op for local dev
 
-from model_inference import load_model, generate_sql
+from model_inference import load_model, generate_sql_streaming
 from data_engine import create_session, execute_safe, QueryTimeoutError
 
 # ── Startup ───────────────────────────────────────────────────────────
@@ -60,32 +62,39 @@ EXAMPLE_QUERIES = [
 @_gpu_decorator
 def handle_query(user_question: str):
     """
-    Process an admin's question end-to-end.
+    Process an admin's question end-to-end, streaming SQL as it generates.
 
-    1. Generate SQL via local LLM (blocking)
+    1. Stream SQL tokens from the LLM into the SQL panel (live)
     2. Execute validated SQL on a fresh per-request DB
-    3. Return (sql_text, dataframe, status_message)
+    3. Yield (sql_text, dataframe, status_message)
     """
     if not user_question or not user_question.strip():
-        return "", None, "⚠️  Please enter a question."
+        yield "", None, "⚠️  Please enter a question."
+        return
 
+    raw_output = ""
     try:
-        raw_output, _ = generate_sql(user_question, llm=llm)
+        yield "", None, "🤖 Generating SQL…"
+        for accumulated in generate_sql_streaming(user_question, llm=llm, max_tokens=192):
+            raw_output = accumulated
+            yield raw_output, None, "🤖 Generating SQL…"
     except Exception as e:
-        return "", None, f"❌ Model error: {e}"
+        yield raw_output, None, f"❌ Model error: {e}"
+        return
 
     try:
+        yield raw_output, None, "🦆 Running query…"
         conn = create_session()
         clean_sql, df = execute_safe(conn, raw_output, timeout_sec=30)
         conn.close()
         row_count = len(df)
-        return clean_sql, df, f"✅ Done — {row_count} row{'s' if row_count != 1 else ''} returned"
+        yield clean_sql, df, f"✅ Done — {row_count} row{'s' if row_count != 1 else ''} returned"
     except ValueError as e:
-        return raw_output, None, f"⚠️ Validation: {e}"
+        yield raw_output, None, f"⚠️ Validation: {e}"
     except QueryTimeoutError as e:
-        return raw_output, None, f"⏱️ Timeout: {e}"
+        yield raw_output, None, f"⏱️ Timeout: {e}"
     except Exception as e:
-        return raw_output, None, f"❌ Error: {e}"
+        yield raw_output, None, f"❌ Error: {e}"
 
 # ── UI ─────────────────────────────────────────────────────────────────
 

prompts.py

@@ -105,14 +105,6 @@ FEW_SHOT_EXAMPLES = [
         "question": "Show total enrollment per school for 2024-2025, sorted highest first.",
         "sql": "SELECT school_name, SUM(student_count) AS total_enrollment\nFROM enrollment\nWHERE school_year = '2024-2025'\nGROUP BY school_name\nORDER BY total_enrollment DESC;",
     },
-    {
-        "question": "What percentage of students at Lincoln Elementary were chronically absent in 2023-2024?",
-        "sql": "SELECT\n  COUNT(CASE WHEN is_chronically_absent THEN 1 END) * 100.0 / COUNT(*) AS chronic_absence_pct\nFROM attendance\nWHERE school_year = '2023-2024' AND school_name = 'Lincoln Elementary';",
-    },
-    {
-        "question": "Show me the enrollment trend for all schools since 2021.",
-        "sql": "SELECT school_year, SUM(student_count) AS total_enrollment\nFROM enrollment\nWHERE school_year >= '2021-2022'\nGROUP BY school_year\nORDER BY school_year;",
-    },
     {
         "question": "How many chronically absent students are English Learners in 2023-2024?",
         "sql": "SELECT COUNT(DISTINCT a.student_id) AS chronic_ell_count\nFROM attendance a\nJOIN students s ON a.student_id = s.student_id\nWHERE a.school_year = '2023-2024'\n  AND a.is_chronically_absent = TRUE\n  AND s.english_learner = TRUE;",
@@ -121,10 +113,6 @@ FEW_SHOT_EXAMPLES = [
         "question": "What's the average GPA for chronically absent students vs non-chronic students in 2023-2024?",
         "sql": "SELECT\n  a.is_chronically_absent,\n  ROUND(AVG(g.gpa), 2) AS avg_gpa\nFROM attendance a\nJOIN grades g ON a.student_id = g.student_id AND a.school_year = g.school_year\nWHERE a.school_year = '2023-2024'\nGROUP BY a.is_chronically_absent;",
     },
-    {
-        "question": "How many discipline incidents were recorded at each school in 2023-2024?",
-        "sql": "SELECT school_name, COUNT(*) AS incident_count\nFROM discipline\nWHERE school_year = '2023-2024'\nGROUP BY school_name\nORDER BY incident_count DESC;",
-    },
 ]