v1.4.6

DeepTutor v1.4.6 Release Notes

Release Date: 2026.06.17

Heads up: the documentation is still catching up to this release, so some
pages may not yet reflect everything below.

A broad consolidation release on top of v1.4.5, organized
around the four surfaces it touches: the Learning Space becomes a real
dashboard, the Knowledge Center gains new retrieval engines and external
mounts, Settings opens up document parsing, voice, and media generation,
and the platform tightens per-user access. Drop-in upgrade — no data migration.

What's New

Learning Space

The Space is now a learning dashboard that pulls your work together, and
the surfaces around it grew up:

• My Agents — import your local Claude Code / Codex chat history into a
  resumable agent you can keep talking to (/space/agents). A referenced
  transcript gets an objective briefing and third-party framing, so the tutor
  narrates it instead of role-playing as you.
• Memory is promoted to its own top-level destination (/memory) instead
  of being buried in the workspace, with the L1/L2/L3 views intact.
• Partner conversations now flow into Memory on the same pipeline as chat.
• Install community skills from EduHub inline at /space/skills.

Knowledge Center

/knowledge is rebuilt as a Knowledge Center console — engine cards are
pure entry points and configuration moves into a per-engine detail page:

• New retrieval engines join LlamaIndex: GraphRAG (microsoft/graphrag),
  PageIndex (cloud reasoning-retrieval), and a built-out LightRAG
  backend — picked per knowledge base.
• Linked KB — point a knowledge base at an already-built index folder as a
  read-only mount, skipping ingestion entirely.
• Obsidian — connect a live vault as an agentic knowledge capability that
  reads and writes notes directly.

Settings

New configurable services, each with its own settings page and all off until
you set them up:

• Document parsing — a pluggable parse layer (MinerU, Docling, MarkItDown)
  with a content-addressed cache (/settings/document-parsing). MinerU adds a
  cloud/local dual backend (/settings/mineru) and now extracts question type,
  difficulty, and answer.
• Voice — a speaker button with three-state autoplay for replies (TTS) and
  a composer mic for dictation (STT), at /settings/tts and /settings/stt.
• Image & video generation — configurable model catalogs wired into chat
  tools (/settings/image, /settings/video).
• Settings navigation rebuilt around the expanded surface.

Other

• Per-user access gating — LLM capabilities are locked for users without an
  assigned model, with the sidebar lock and the server-side gate always in
  agreement.
• Solve is rebuilt as a capability on the chat loop, with a per-step gate
  and a mid-run failure fallback that forces a graceful finish instead of
  failing the whole turn.
• Login accepts a username (not just an email), surfaces bad-login errors
  inline instead of silently reloading, and fixes the NEXT_PUBLIC_AUTH_ENABLED
  flag.
• CLI gains deeptutor skill commands — browser login plus publish/update
  for your own skills.
• Context pickers are polished onto a shared shell with live preview, and a
  math-before-citation rendering fix keeps $…$ from being mangled by citation
  linkification.

Upgrade Notes

• Drop-in from v1.4.5: pip install -U deeptutor; Docker users pull
  ghcr.io/hkuds/deeptutor:latest. No migrations — existing data loads as-is.
• Assign each user a model so their capabilities unlock; admins manage this
  from the model catalog.
• New engines and services are opt-in: GraphRAG needs the extra
  (pip install 'deeptutor[graphrag]'), document parsing needs
  deeptutor[parse], and voice / image / video need their model catalogs
  configured.

Full Changelog: v1.4.5...v1.4.6