OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Timesearch R

TimeSearch-R is an open-source Python research model and implementation for adaptive temporal search in long-form video. It interleaves text and video reasoning with reinforcement learning, and its GRPO-CSV method uses self-verification to assess whether searched frames adequately support video reasoning.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Timesearch R — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when you want a research codebase and model checkpoint for local experiments in long-video temporal search, especially where video QA depends on finding relevant moments rather than processing a whole video as one input.

Who it is for

  • Researchers studying long-form video understanding and temporal search.
  • Developers building experimental video question-answering workflows that need to retrieve supporting clips.
  • Teams evaluating model behavior on the documented Haystack-LVBench and LongVideoBench data.

Documented uses

  • Search long videos for temporally relevant clips.
  • Retrieve video moments that support a question-answering workflow.
  • Evaluate temporal-search and long-video QA results with the repository's scripts.
  • Study or train a self-verifying temporal-search approach.
Before you choose

Requirements & limits

The public repository identifies the project as open source under the Apache 2.0 license and links to a model checkpoint.

  • The documented implementation is Python-based and expects CUDA and PyTorch.
  • Inference uses a separate SigLIP clip server for video-frame retrieval.
  • The repository recommends preprocessing frame and feature caches to accelerate inference and training.
  • Training has substantial stated resource recommendations: at least 16 GPUs and 1,024 GB CPU RAM per node for 7B training.
  • The documented workflow uses Haystack-LVBench JSON files and corresponding LongVideoBench videos.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run documented inference on Haystack-LVBench

Prepare the documented model, data and clip-retrieval service for the repository's inference workflow.

Before you start

  • CUDA and PyTorch; the repository reports CUDA 12.4 and PyTorch 2.6.0 in its experiments.
  • Python dependencies installed with pip from requirements.txt.
  • A downloaded SigLIP model, with its local path set in the clip-server configuration.
  • The TimeSearch-R checkpoint, Haystack-LVBench JSON input and corresponding LongVideoBench MP4 files.

Steps

  1. Install the repository dependencies with pip install -r requirements.txt.

  2. Download SigLIP, set its local path in clip_as_service/server/clip_server/torch-flow.yml, then install and start the clip server as documented.

  3. Download the TimeSearch-R checkpoint and prepare the dataset files; optionally create the recommended frame and feature caches.

  4. Set SIGLIP_URL to the clip server address and run time_r1/inference.py with torchrun, supplying the input, output, data-root and checkpoint paths.

Expected outcome

The documented workflow writes inference results to the selected output path; the repository also provides separate temporal-search and question-answering evaluation scripts.

Guide limitations

  • The example uses eight processes and requires local GPU infrastructure.
  • Frame and feature caching is recommended for speed, not listed as a mandatory setup step.
  • This guide describes repository instructions, not a verified execution or compatibility result.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
Time-Search/TimeSearch-R · Official homepage title · retained in its source language
How we review sourcesSuggest a correction