Skip to content

v0.4.0 — 2026-05-18

Choose a tag to compare

View all tags
@nduckmink nduckmink released this 18 May 14:38
v0.4.0
57a9c7b

Changelog

All notable changes to Arkon are documented here.
Format follows Keep a Changelog.

[0.4.0] — 2026-05-18

A scope-aware refresh of the Wiki UX. Every wiki page already lived in a
scope (global, department, project), but the UI treated same-slug
pages from different scopes as duplicates and the index page only ever
showed the global catalog. This release threads scope context through
the URL, the page tree, the page detail view, and the ingestion pipeline.

Added

  • Scope switcher on /wiki: a dropdown listing the scopes the
    current user can access. Selecting a scope updates the URL
    (/wiki?scope_type=&scope_id=), refetches the matching _index
    catalog and pages grid, and is shareable / reloadable.
  • Scope-grouped page tree: the sidebar in /wiki and the detail
    viewer now groups pages by scope (GLOBAL, plus each department),
    then by page type (Entities, Concepts, Topics, Sources).
    Clicking a scope header opens that scope's wiki landing; the chevron
    toggles expansion independently. The active scope and the bucket
    containing the active page auto-expand on navigation.
  • Scope-preserving navigation: backlinks, outlinks, and inline
    [[wikilinks]] carry the current page's scope params, so jumping
    between related pages keeps the user inside the same scope context.
    The detail back button returns to /wiki?scope_type=&scope_id= for
    department-scoped pages (projects continue to return to
    /workspaces).
  • GET /api/wiki/my-scopes: lists global plus each department and
    project the requester can read. Used by the scope switcher.
  • scope_name in wiki responses: /api/wiki/pages joins the
    Project and Department tables so each summary carries a
    human-readable scope name (e.g. "Phòng Nhân sự") alongside the raw
    ID, removing the need for separate lookups on the client.
  • Thematic-section concept pages during ingestion: the MRP
    extraction and planning prompts now recognise documents that describe
    a primary entity through several distinct themes (e.g. Product
    Positioning    , Target Customer Profile, Content Pillars) and emit
    a separate concept page per theme instead of dumping the content
    into the entity page. The entity page links out to them with
    [[concept/...]].

Changed

  • /api/wiki/pages and /api/wiki/index accept optional
    scope_type + scope_id query parameters. When omitted the original
    behaviour is preserved (RBAC-filtered list, global index).
  • ScopeBadge accepts scopeId: string | null to match the relaxed
    WikiPageSummary type.
  • Sidebar collapse/expand chevrons swapped for left_panel_close /
    left_panel_open icons that don't look like a Back button.
  • The /wiki page tree no longer surfaces project-scoped pages —
    workspaces remain reachable from /workspaces and from the scope
    switcher, keeping the wiki sidebar focused on enterprise-wide
    knowledge.
  • The wiki graph no longer draws the dashed convex-hull boundaries
    around department and project clusters; nodes and edges stand on
    their own.
  • The Wiki button previously added to department cards was removed
    once the scope switcher and clickable tree scope headers landed —
    redundant entry points were creating clutter.

Fixed

  • Department-scoped detail pages no longer fetch
    /api/projects/<id>/wiki (which 404s on department IDs); the tree
    uses the new general scope-aware /api/wiki/pages endpoint instead.
  • Clicking a backlink, outlink, or inline [[wikilink]] from a scoped
    page used to drop scope context and load the flat "old" tree.
  • Navigating to a global wiki page used to render with the legacy flat
    tree while /wiki showed the new grouped tree.
  • DELETE /api/wiki/pages/<slug> returned 404 for workspace pages
    because the endpoint looked them up with the default global scope.
    Even after that lookup was fixed, the cascade helper re-fetched the
    row with the same default and silently no-op'd the actual delete,
    returning {"ok": true} while the row remained in the database.
    delete_page_cascade now takes the resolved WikiPage object so
    no second lookup is performed.

Backend

  • Frontend WikiPageSummary and Pydantic WikiPageSummary both gain
    scope_name: Optional[str]; new shared WikiScope type for the
    switcher payload.
  • _build_wiki_scope_filter is reused by both /wiki/pages and the
    new /wiki/my-scopes endpoint.
  • regenerate_index is called with the deleted page's actual scope so
    the right _index is rebuilt after a non-global delete.

[0.3.1] — 2026-05-14

Added

  • Wiki Graph — Department Clustering: Wiki pages scoped to a department now visually group into department clusters on the /wiki/graph canvas.
    • Convex hull drawn per department (below project hulls) with a distinct color per department.
    • Force simulation biases nodes toward their department's X-zone (70% scope pull, 30% component spread) so related pages naturally converge.
    • Legend lists each department with icon business and page count.
    • Tooltip shows department name for dept-scoped pages.

Fixed

  • Graph endpoint now joins the Department table so scope_name is populated for department-scoped pages (previously only Project was joined, leaving dept nodes without a name label).

[0.3.0] — 2026-05-13

Added

  • Department-level Wiki Isolation: Wiki pages compiled from department-scoped sources are now restricted to members of that department.

    • ScopeType.DEPARTMENT added to the enum; pipeline _resolve_wiki_scopes() resolves project > department(s) > global, fanning out multi-department sources into one page per department scope (LLM runs once, content is duplicated to each scope).
    • wiki_service._scope_filter_with_dept() provides a single-query OR filter (global + user's department).
    • get_wiki_page returns HTTP 403 for cross-department access.
    • Source PATCH: changing department on a ready source triggers wiki detach, old-scope index regeneration, and MRP re-queue automatically.
    • Frontend: edit-source dialog warns before department reassignment triggers re-analysis.

  • MRP Pipeline — Plan Regeneration with Reviewer Feedback: Admin can now reject a pending plan with a note, triggering LLM-based regeneration that incorporates the feedback.

    • POST /sources/{id}/plan/regenerate runs in the background via regenerate_plan_task.
    • Plan Review Dialog surfaces a Regenerate button that requires a reviewer note.
    • _resolve_maybe_items uses LLM to decide UPDATE vs CREATE (previously always downgraded MAYBE to CREATE).

  • Catalog-driven LLM & Vision Selection: Replaces free-form llm_provider + llm_model_id config with curated catalogs (LLMModelSpec, VisionModelSpec) that expose context window size, tool support, vision capability, and per-token cost.

    • /api/settings/{llm,vision}/{catalog,switch} endpoints mirror the embedding catalog pattern.
    • Settings UI renders a ModelCatalogCard per capability with metadata (context window, costs, tool/vision badges).
    • writer._get_source_context_budget reads context_window_tokens from the spec — the stale hard-coded table is removed.

  • Gemini Model Updates: Catalog updated with newer Gemini variants.

    • gemini-3.1-flash-lite: 1M context, tools + vision + thinking, cheapest Google 1M option ($0.25 in / $1.50 out per 1M tokens). New recommended default for high-volume extraction and captioning.
    • gemini-3-flash-preview and two additional preview models added.
    • Admins on gemini-3.1-flash must reselect in Settings (model removed from catalog).

Fixed

  • MRP Pipeline Hardening (critical):

    • Draft results (PageWriteResult) now persisted in plan_json._page_drafts; VERIFY/COMMIT phases resume without re-running REFINE.
    • caption_images_task is now serialized before ingest_map_reduce_task, baking captions into source.full_text before MAP runs — fixes the race condition that produced empty image markers in compiled wiki pages.
    • KB reconciliation searches every destination scope and retains the best semantic match, preventing duplicate pages when the same concept exists across scopes.

  • MRP Pipeline Hardening (high):

    • assemble_evidence uses word-boundary regex (\bterm\b) instead of substring matching, so short entity names (e.g. "AI") no longer match unrelated subjects ("MAIL").
    • /sources/{id}/plan/regenerate runs async via arq; UI polls GET /plan instead of holding an open HTTP connection.
    • JSON fence stripping unified via parse_json_loose; removes several incorrect str.strip("```json") variants in mapper and wiki_analyzer.

  • MRP Pipeline Hardening (medium):

    • Approve/reject/regenerate endpoints use SELECT FOR UPDATE and reject mismatched status to prevent race conditions.
Assets 2
Loading