This release focuses on significantly improving the robustness and reliability of LLM response parsing. By implementing a multi-layered JSON extraction strategy, the Deep Dive engine can now handle non-compliant or messy outputs from a wider range of models (including smaller or less-steered models) without failing the analysis chain.
New Features
- 🛡️ Robust JSON Extraction Engine:
- Introduced a new
_extract_jsonmethod that prioritizes Markdown code block detection (supports```jsonand raw```blocks). - Implemented a "Greedy Brace Matching" fallback that scans the entire output for the first
{and last}to rescue valid JSON buried in conversational filler.
- Introduced a new
- 🧹 Automatic Syntax Cleaning:
- Added a safety layer that automatically identifies and removes common JSON syntax errors, such as trailing commas in objects or arrays, before parsing.
Improvements
- 🏗️ Architecture Refactoring:
- Decoupled the parsing and extraction logic from the main
actionflow. - Improved internal logging for easier debugging of LLM formatting issues.
- Decoupled the parsing and extraction logic from the main
- 🌐 Version Synchronization: All display versions across the UI and documentation have been aligned to v1.1.0.
Bug Fixes
- Fixed a potential crash when LLMs provided multiple JSON blocks or included explanatory text outside of the JSON structure.
- Resolved an issue where minor JSON syntax errors would cause the entire "Deep Dive" process to fail silently.
Version Changes
Plugin Updates
- Deep Dive: v1.0.1 → v1.1.0 | 📖 README