DeepTutor v1.0.3 Release Notes
Release Date: 2026.04.13
Highlights
Question Notebook — Unified Quiz Review System
Replaced the single-purpose "Wrong Answer Note" with a full Question Notebook that stores every quiz question (correct and incorrect) with rich metadata — question type, options, explanation, and difficulty. Each entry supports bookmarking and category tagging for organised review. A new dedicated /notebook page provides filtering (all / bookmarked / wrong), category management (create, rename, delete), and direct links back to the originating session. The QuizViewer component now integrates inline bookmark and category controls so users can organise questions without leaving the quiz flow.
Mermaid Diagram Support in Visualize
Extended the Visualize capability with a third render type — Mermaid. The analysis agent now chooses between svg, chartjs, and mermaid, preferring Mermaid for structured diagrams (flowcharts, sequence diagrams, class diagrams, mindmaps, etc.). The code generator and review agents received corresponding prompt updates, and the frontend VisualizationViewer renders Mermaid diagrams via a dedicated <Mermaid> component.
Embedding Model Mismatch Detection
Knowledge bases now record the embedding model and dimension used at index time. On load, the system compares stored fingerprints against the currently configured embedding model; mismatched KBs are flagged with embedding_mismatch and needs_reindex. The RAG search pipeline surfaces a warning when querying a mismatched KB, and the Knowledge page displays an alert badge so users know to re-index.
System Message Merging for Qwen / vLLM Compatibility
Consolidated multiple system messages into a single merged system message in both AgenticChatPipeline and ChatAgent, fixing compatibility with Qwen models served via vLLM that reject multi-system-message conversations. The context builder now filters duplicate system messages from stored history to avoid redundancy.
LM Studio & llama.cpp Provider Support
Added first-class ProviderSpec entries for LM Studio (localhost:1234) and llama.cpp (localhost:8080) with automatic base-URL detection and the openai_compat backend, plus embedding-provider alias mapping.
Glass Theme
Introduced a new Glass theme with frosted-glass card surfaces, gradient backgrounds, and glow-accent buttons. The theme switcher now cycles through light → dark → glass.
Deep Research Reporting Agent Resilience
Extracted a shared _call_llm_json helper with configurable retry logic in the reporting agent, replacing three identical inline LLM-call-then-parse blocks for introduction, section body, and conclusion generation.
Documentation Migration
Removed the legacy VitePress docs/ folder; documentation has been migrated to the project website.
Community Contributions
- @cskwork — Wrong-answer note for cross-session quiz review (#292)
- @OlegSob-glitch — Merge system messages and add history reference fallback (#295)
- @SuperMarioYL — Embedding model mismatch detection in knowledge bases (#299)
Full Changelog: v1.0.2...v1.0.3