---
title: ToolMerge Video Keyframe Retrieval
emoji: 🎬
colorFrom: pink
colorTo: gray
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
short_description: Decompose queries into tool calls for keyframe retrieval
python_version: "3.12"
startup_duration_timeout: 1h
pinned: false
---

# ToolMerge — Long-Video Keyframe Retrieval

Live demo of **[ToolMerge](https://github.com/michalsr/ToolMerge)**
("Decomposing Queries into Tool Calls for Long-Video Keyframe Retrieval",
[paper 2605.23826](https://huggingface.co/papers/2605.23826)).

A text-only **Qwen3-VL planner** ([`michalsr/toolmerge-planner-grpo`](https://huggingface.co/michalsr/toolmerge-planner-grpo))
decomposes your question + answer choices into independent visual-search
tool calls combined with **AND / OR** boolean operators. Each query is
text-encoded with [SigLIP-2](https://huggingface.co/google/siglip2-giant-opt-patch16-384)
and scored against every sampled frame; the per-tool percentile ranks are
merged (AND = min, OR = max), and greedy temporal NMS
(τ = min(D/2K, 10)s) selects the final top-K keyframes.

This demo runs the paper's **SigLIP path**. The T-REN region-text tool and the
OCR-judge stage require offline-built per-video caches and separate weights and
are omitted; the boolean merge runs SigLIP-only, exactly as the reference code
does when other tools are disabled via `enabled_tools`.