Starstruk Logo MarkStarstruk

v1.12.2

bytedance/flowgram.aiv1.12.2Apr 10, 2026by ItzCrazyKns

AI Summary

This release focuses on stabilizing the research pipeline and significantly enhancing Deep Research capabilities. It introduces a new Chromium-based scraper, optimizes search execution using embeddings, and improves error handling to prevent crashes.

Key Highlights

  • Deep Research Mode now uses an aggressive Reason-Search-Scrape-Extract-Repeat cycle with dynamic context window management to prevent overflow.
  • Integrated a Chromium-based scraper for better compatibility with modern web pages.
  • Implemented a new, optimized `executeSearch` function leveraging embeddings to filter results.
  • Improved widget execution handling to ensure errors do not crash the entire pipeline.
  • Added validation and timeouts to prevent hung search requests.

New Features

  • Chromium-based scraper
  • Optimized `executeSearch` function with embeddings
  • Dynamic context window management for Deep Research
  • Integrated `serverUtils` and updated reverse geolocation API

Full Release Notes

## What's Changed

  * Fixed history double-conversion bugs to ensure smoother suggestion generation.
  * Added validation and timeouts to prevent hung search requests.
  * Improved widget execution handling; errors no longer crash the entire research pipeline.
  * Resolved build-time errors by addressing missing package dependencies.
  * Fixed edge cases involving non-array search query formats.
  * Integrated `serverUtils` and updated the reverse geolocation API for better accuracy.
  * Enhanced error handling for file upload workflows.
  * Added a Chromium-based scraper for better compatibility with modern web pages.
  * Implemented a new, optimized `executeSearch` function.
  * Leveraged embeddings to filter search results, preventing context window overflow and improving relevance.

## Improvements in Deep Research Mode

The Deep Research pipeline now operates on a more aggressive Reason-Search-Scrape-Extract-Repeat cycle. The system iteratively identifies top-tier websites, scrapes their full content, and extracts key data points to inform subsequent research steps before generating the final response. While this adds a bit more processing time, the depth of insight is substantially increased.

To support this high volume data extraction, we now manage the context window dynamically. By processing and feeding data in optimized chunks, the system prevents context overflow while ensuring all relevant information from the multi-stage research process is considered.

![https://github.com/user-attachments/assets/bc7bed6a-2cae-4a47-b39a-bcbd6d57ba40](https://github.com/user-attachments/assets/bc7bed6a-2cae-4a47-b39a-bcbd6d57ba40)

## New Contributors

  * @marexxxxxxx made their first contribution in [#1041](https://github.com/ItzCrazyKns/Vane/pull/1041)
  * @joaquinescalante23 made their first contribution in [#1015](https://github.com/ItzCrazyKns/Vane/pull/1015)
  * @saschabuehrle made their first contribution in [#1076](https://github.com/ItzCrazyKns/Vane/pull/1076)
  * @nickorlabs made their first contribution in [#1082](https://github.com/ItzCrazyKns/Vane/pull/1082)

**Full Changelog**: [v1.12.1...v1.12.2](https://github.com/ItzCrazyKns/Vane/compare/v1.12.1...v1.12.2)