Neura Market
prompt logo

prompt

Free

Sustained multimodal search agent with horizon up to 100 turns

FreeFree tier
Type
Open Source
Company
ai-boost
Visit Website GitHub (40 stars)

About prompt

The Long-Horizon Multimodal Search Agent is an advanced system prompt designed for AI agents to perform sustained, multi-turn visual and textual searches across up to 100 search steps. It implements file-based visual context management, progressive on-demand image loading, and search trajectory planning to avoid common failure modes such as context bloat, visual memory loss, redundant re-search, hallucinated visual claims, and horizon collapse. The prompt structures the agent's reasoning into core responsibilities including maintaining a visual index with UID references, loading images only when required, batching visual requests, and conducting horizon reviews every 10 turns to re-plan based on evidence gathered. It supports multi-hop visual reasoning with three hops: locating candidate sources, extracting visual evidence, and synthesizing findings. The agent is tested on benchmarks like MM-BrowseComp and MMSearch-Plus, and scales to 100-turn search horizons.

Key Features

File-based visual context management with unique image UIDs and per-image metadata (source URL, load turn, thumbnail summary, confidence)
Progressive on-demand image loading: thumbnail-first, full-resolution only when needed, batch loading, and keyframe sampling for video
Search trajectory planning with a search tree, priority assignments, depth budgets, and horizon reviews every 10 turns
Multi-hop visual reasoning across three hops: locate candidates, extract visual evidence, synthesize findings
Visual index maintenance to avoid redundant re-search and visual memory loss
Handles up to 100 search steps without context bloat or token exhaustion
State-of-the-art performance on MM-BrowseComp and MMSearch-Plus benchmarks

Pros & Cons

Pros
  • Eliminates context bloat by offloading full-resolution images and keeping only UID references and compressed summaries
  • Prevents visual memory loss with a persistent visual index and horizon review mechanism
  • Avoids redundant re-search through structured search trajectory planning
  • Reduces hallucination by only describing images that were actually loaded and recorded
  • Scales to 100-turn search horizons without horizon collapse
  • Achieves state-of-the-art results on multimodal search benchmarks
Cons
  • Requires a system capable of managing file-based visual context and executing complex trajectory planning
  • May have higher upfront planning overhead due to search tree drafting and horizon reviews
  • Dependent on the underlying model's ability to follow a multi-step structured prompt accurately
  • Performance may degrade if the model fails to adhere to the indexing and loading rules

Best For

Complex information gathering tasks requiring sustained visual and textual search across many sourcesComparative analysis of images from multiple web pages or documents over multiple turnsLong-horizon research where the agent must track what has been seen and plan future search stepsMultimodal fact-checking that requires loading and analyzing images progressivelyVideo content analysis using keyframe sampling to process video evidence in a search trajectory

FAQ

What is the maximum number of search steps this agent can handle?
The agent is designed to handle up to 100 search turns without context bloat or horizon collapse.
How does the agent avoid visual hallucination?
It only loads images when required by the current reasoning step, records a visual index with source URLs and thumbnails, and never describes images that were not actually loaded.
What benchmarks has this agent been tested on?
It achieves state-of-the-art performance on MM-BrowseComp and MMSearch-Plus benchmarks.
Is the agent free to use?
The prompt is open source and available for free in the ai-boost/awesome-prompts repository on GitHub.