- Without Ratel: drop the top-k list; show the direct tool call over the full catalog
- With Ratel: reveal retrieval before invoke_tool, since the model picks the tool
- Slower baseline generation latency to reflect the larger context
- Score recall as gold coverage of each side's context; drop precision
* Replace native prompt select with a themed dark dropdown
The native <select> popup used the OS's light styling, clashing with the dark
theme. Swap it for a custom listbox: dark forest menu, cream options, coral
focus ring, hover/active highlighting, and keyboard support (arrows/Enter/Escape).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Prompt dropdown fills from the retriever toggle to the right edge
Flex control row: the retriever toggle keeps its width, the dropdown starts
right after it and grows to the layout edge.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Step-by-step gateway traces, 100-tool catalog, Run button, per-panel metrics
Expand the catalog to 100 tools (rebaked retrieval + real Qwen3-4b runs) and
rework the compare UI:
- Both panels play a step-by-step trace: Without Ratel runs the native-MCP
baseline (search_operations -> execute_operation), With Ratel runs the
gateway (search_capabilities -> invoke_t…
Redesign demo UI: dark ratel.sh theme, chat-style runs, top-3, more prompts (#2)
Berca Akbayır+1: Claude Opus 4.8 (1M context)
Unverified signature
e09b1e2
Rework the Space to match ratel.sh: dark forest background (#001c16) with
lighter forest surfaces, cream text, brand green + orange accents, a pixel-font
"Ratel — Demo" title, and simplified copy. The model runs now render as a
kestral-style chat (user bubble → search_tools → tool call → streaming answer),
with the retrieved tools listed after the answer finishes. Retrieval is top-3;
prompts expanded to 14, re-baked over qwen3-4b; tool selections are colored
hit/partial/miss.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add Ratel "with vs without" tool-retrieval Hugging Face Space (#1)
Berca Akbayır+1: Claude Opus 4.8 (1M context)
Unverified signature
6e72953
Static Space that replays real, offline-baked runs: Ratel retrieves the top-5
tools from a fixed 70-tool catalog vs loading all 70 into context, and each
side shows a saved qwen3-4b run (selected tools, answer, measured token cost,
recall) per retriever method (sparse/dense/hybrid). Retrieval baked with the
real Ratel SDK (bake.py); model runs baked with bake_llm.py.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>