tune+ux: z-thresh 1.8 + new examples + 'development practices' copy (app.py)

app.py

@@ -15,23 +15,6 @@ proportional to actual model work.
 from __future__ import annotations
 
 import os
-
-# `import spaces` MUST come before any module that imports torch — the
-# spaces package installs a torch CUDA-emulation patch at import time,
-# and if torch has already loaded the patch can't take. Symptom of
-# getting this wrong on ZeroGPU: worker_init fails with
-#   RuntimeError: No CUDA GPUs are available
-# at torch.init(nvidia_uuid). Keep this block above the imports below.
-#
-# Set USE_SPACES_GPU=false on persistent paid-GPU Spaces (h200,
-# a10g-small, etc.) where the decorator isn't needed.
-USE_SPACES_GPU = os.environ.get("USE_SPACES_GPU", "true").lower() in ("true", "1", "yes")
-if USE_SPACES_GPU:
-    import spaces
-    _gpu_decorator = spaces.GPU(duration=60)
-else:
-    _gpu_decorator = lambda f: f
-
 import time
 import traceback
 from typing import Any
@@ -45,6 +28,21 @@ import confidence
 import model_io
 import team_context
 
+# The @spaces.GPU decorator is for Zero-GPU Spaces only — it spawns a
+# worker process that requests a GPU from the Zero-GPU pool. On
+# persistent paid GPU Spaces (h200, a10g-small, etc.), the GPU is
+# always attached to the main container; the decorator would either be
+# a no-op or fail at worker-init.
+#
+# Set USE_SPACES_GPU=false on persistent GPU Spaces. Default true so
+# Zero-GPU deployments keep working unchanged.
+USE_SPACES_GPU = os.environ.get("USE_SPACES_GPU", "true").lower() in ("true", "1", "yes")
+if USE_SPACES_GPU:
+    import spaces
+    _gpu_decorator = spaces.GPU(duration=60)
+else:
+    _gpu_decorator = lambda f: f   # no-op on persistent GPU
+
 # Loaded at cold-start (CPU). The GPU model is loaded inside the
 # scoring function on first invocation.
 EMBEDDER_PRELOAD = candidate_filter.load_embedder()
@@ -157,18 +155,17 @@ def _format_result(result: dict, team_ctx: str, elapsed_s: float) -> tuple[str,
         # generalist; don't push hard, but keep the upsell visible.
         confidence_md += (
             "\n\n---\n\n"
-            "_Want personalized recommendations that learn your team's "
-            "specific reading patterns?_\n\n"
+            "_Want recommendations tailored to your team's development practices?_\n\n"
             f"{cta_button}"
         )
     else:
         # Prominent CTA — user saw the generalist's limits (moderate /
         # low / noise tier). Personalized scoring directly addresses the
-        # cause: a model adapted to their team's reading history.
+        # cause: a model adapted to the team's development practices.
         confidence_md += (
             "\n\n---\n\n"
             "### Get sharper recommendations on this repo\n\n"
-            "A personalized model trained on your team's full reading history "
+            "A personalized model tailored to your team's development practices "
             "typically lifts confidence by **10-15pts** on team-specific picks — "
             "and produces PR drafts that match your team's actual engineering style.\n\n"
             f"{cta_button}"
@@ -281,8 +278,8 @@ with gr.Blocks(title="Feature Finder", theme="soft") as demo:
         "a PR-ready spec (summary, motivation, implementation plan, open "
         "questions) that a coding agent can pick up and run with.\n\n"
         "Free preview using our open-source generalist scorer (LoRA on "
-        "Qwen3.5-2B). For personalized recommendations trained on your team's "
-        "full reading history: "
+        "Qwen3.5-2B). For recommendations tailored to your team's "
+        "development practices: "
         '<a href="https://engine.remyx.ai" target="_blank" '
         'style="color:#ff6b35;font-weight:600;text-decoration:underline;">'
         "sign up at engine.remyx.ai</a>.",
@@ -315,17 +312,20 @@ with gr.Blocks(title="Feature Finder", theme="soft") as demo:
 
     gr.Examples(
         examples=[
-            # Remyx's own — multimodal data synthesis; demos the moderate-tier
-            # "topically adjacent but not perfect" case + the CTA value-prop.
+            # Remyx's own — multimodal data synthesis; demos the moderate
+            # tier on a real internal repo + the CTA value-prop.
             "https://github.com/remyxai/VQASynth",
-            # Agent framework with clear domain — typically lands high tier.
-            "https://github.com/AgentEra/Agently",
+            # Popular multi-agent framework — recognizable to most ML devs,
+            # typically lands high tier with STORM-style multi-agent papers.
+            "https://github.com/microsoft/autogen",
+            # Popular RAG / data-indexing framework — strong high-tier
+            # demo case; relevant retrieval-augmented papers are dense
+            # in the pool.
+            "https://github.com/run-llama/llama_index",
             # Hyperparameter optimization library — less mainstream, tests
-            # the LLM-training/ML adjacency.
+            # the LLM-training/ML adjacency and demos honest low-tier
+            # behavior on niche subdomains.
             "https://github.com/hyperactive-project/Hyperactive",
-            # Rust + React coding assistant — niche but in active development;
-            # good cross-domain test.
-            "https://github.com/amrit110/oli",
         ],
         inputs=[repo_input],
     )