v0.55.0 — Sandbox probes through an executor sidecar

[0.55.0] - 2026-10-21

The sandbox console runs its connection probe through an executor

• The admin sandbox page's Test button now runs the upstream probe for the
  E2B and Aliyun Code Interpreter providers through a deployer-provided
  executor sidecar (RAYRAG_SANDBOX_EXECUTOR_URL): the provider config and the
  upstream test snippet are POSTed to that service and the answer is reported
  with the upstream success rule and message format.
• Without an executor the endpoint says the provider runs through one instead
  of faking a pass; missing credentials are reported before any network call.
  This matches v0.27.2, where the E2B provider's own execution is still a TODO
  and the Aliyun provider depends on the proprietary agentrun-sdk, which the
  image does not bundle.
• Unit tests cover the sidecar request/response and both error paths; the
  sandbox admin endpoint test and a CDP probe (E2B and Aliyun selected in the
  console) verify the full flow.

v0.54.0 — Vendored zvec 0.7.2 and the shared model tree closes

[0.54.0] - 2026-10-21

The Docker build links the vendored zvec 0.7.2 crates

• The vendored zvec Rust crates were still 0.7.1 while the Cargo tag resolves
  0.7.2, so the Docker [patch] no longer matched and the build silently fell
  back to fetching the git dependency (the patch was not used warning). The
  vendor tree now holds the v0.7.2 crates, VENDOR.md records the new tag, and
  ARG ZVEC_RUST_VERSION points at the matching prebuilt release, so the patch
  applies and the build stays self-contained.

The shared model tree closes

• All in-scope consumers of the shared provider/instance/model selector are
  covered: dataset create and configuration, retrieval-test rerank, memory
  list and detail embedding/LLM, skills search configuration, the search page
  rerank field, the search-app detail rerank field and the six default-model
  fields. The remaining upstream consumer files are reference material by
  project scope (RayRAG serves a Rust-native management surface).

v0.53.0 — Real OAuth tokens for the data-source fields

[0.53.0] - 2026-10-21

The data-source OAuth token fields mint real tokens

• The Box and Gmail/Google-Drive token fields exchange the authorization code at
  the configured token endpoint (BOX_OAUTH_TOKEN_URL / GOOGLE_OAUTH_TOKEN_URL)
  and store the returned access and refresh tokens in the connector credentials.
  The exchange is extracted into a shared helper with unit tests that post the
  authorization-code form to a local one-shot endpoint and assert the stored
  tokens, plus the no-endpoint path that keeps the code.
• End-to-end probe: Box start → popup callback → result returns the minted
  tokens; Gmail does the same and the resulting credentials create a gmail
  connector; the mock token endpoint logs both authorization-code form posts.

v0.52.0 — The search answer stream and the mind map

[0.52.0] - 2026-10-21

The search answer stream and the mind map work end to end

• Fixed a real streaming panic: the search app completion bridge sent chunks with
  blocking_send on a bounded channel from inside the async runtime, which
  panicked on the first streamed chunk and left the AI-summary block empty. It
  now uses an unbounded channel; the probe renders the streamed answer, one
  citation chip and the divider.
• Fixed a page-level ReferenceError: searchRerankToggle is not defined when the
  search page is opened with a saved app (?sapp=): the seed applier ran before
  the later script defines the toggle. The call is now guarded.
• Verified the mind-map flow against a reachable chat model: the button opens
  the sheet, the mock model's tree renders seven nodes, and close/clear behave.
• Serialized the OAuth flow-registry unit tests: two tests cleared the process-wide
  registry concurrently and could flake the round-trip assertion.
• Withdrawn a stale residual: the upstream v0.27.2 tree-select rows are plain
  divs with hover/selected accent styles and no Radix focus ring, so the old
  "approximated focus styling" note no longer applies; RayRAG's keyboard roving
  focus is an accessibility superset.

v0.51.0 — Settings shell, data-source grid and search payload

[0.51.0] - 2026-10-21

Search settings edits reach the retrieval request

• Discovered and fixed a real payload bug: the settings drawer marked the
  similarity-threshold and vector-weight sliders with their element ids
  (thresh/vWeight) while the retrieval builder checked threshold/weight,
  so edited slider values never reached the request. ssTouch now records the
  canonical keys too; the probe confirms an untouched request relies on the
  saved search app config and an edited request carries the new threshold.
• The user-setting shell now uses the upstream geometry exactly: a 303 px
  sidebar from the md breakpoint and 64 px below it, with the active item
  derived from the route and kept in sync across back/forward navigation.
• The data-source grid uses the upstream breakpoint set (1 column, 2 columns
  from 1024 px, 4 columns from 1536 px) instead of container queries that never
  reached the 4-column layout; the upstream page has no per-source search box,
  so the old residual is withdrawn.
• Re-audited the UI route table: all /chunk views, the chat widget embed and
  the agentbots log endpoint are registered and reachable, so the routes row is
  aligned.

v0.50.0 — Provider image assets and the UI audit

[0.50.0] - 2026-10-21

Fixed provider image assets and the remaining UI audit

• Every upstream SVG-backed model-provider icon now ships with RayRAG: the 29
  fixed web/src/assets/svg/llm/*.svg assets from v0.27.2 are vendored into the
  static tree, and provider_icon resolves the AIMLAPI, Astraflow, FastEmbed,
  FunASR, FuturMix, GreenPT, MWS, Synthorai and Xiaomi factories to them. A unit
  test now asserts each mapped asset exists.
• The profile avatar uploader is re-verified through the shared crop dialog: the
  profile CDP probe picks an image, confirms the 64x64 crop, checks the PATCH
  round-trip, removes the avatar, and checks both upstream name-validation
  messages and the zh copy.
• The UI re-audit batch re-ran the login-next backgrounds/form, skills search
  configuration, chat channels, model-provider settings, dataset retrieval
  testing and profile pages on the release build; all passed with zero page
  errors, and their ledger rows are aligned.

v0.49.0 — Shared avatar cropper for search settings

[0.49.0] - 2026-10-21

The search settings avatar uses the shared cropper

• The search page's settings drawer now uploads its avatar through the same
  crop dialog as the user profile: pick an image, drag/zoom the square in the
  shared avatar-upload modal, and the 64x64 data URL is shown in the drawer
  with a remove button. Non-image files and images over 4 MB are rejected with
  the upstream copy.
• SearchAppSeed gained the avatar column (POST/PUT /api/v1/searchapps),
  so the cropped image survives save and reload exactly like the React
  AvatarNameDescription field.
• The avatar crop dialog is now mounted once in the shared page shell instead of
  only inside the user-settings area, so every avatar uploader opens the same
  modal.

Admin console re-audited against v0.27.2

• The admin console CDP battery was re-run on the release build and all
  reachable pages matched: console login/guard/cache trio, the services table
  (search, paging, hover action icons, typed JSON view and detail dialogs), the
  users table (superuser selector, password/delete dialogs, own-row hiding) and
  the user detail page with its datasets/agents tables.

v0.48.0 — The Artifacts tab shows the compiled wiki

[0.48.0] - 2026-10-20

The Artifacts tab shows the compiled wiki

The dataset Artifacts tab is now the full RAGFlow v0.27.2 knowledge-compilation
surface instead of a shell around a generate button.

• Every view mode has its upstream empty state and title: Wiki, Graph, Mind map,
  Timeline and the navigation tree. The Generate button follows the upstream
  canGenerateWiki gate, and the structure views explain that their artifacts
  are compiled automatically when documents are parsed.
• The wiki view renders the contents panel (topics -> pages, search, graph tab),
  Update and Clear actions, and a markdown detail pane with headings, lists,
  code and clickable [[wiki links]], plus the source-document line.
• Generation runs end to end from the UI: the MAP extracts are persisted, COMPILE
  synthesises the pages and the view reloads with them. A freshly parsed dataset
  now compiles in one pass: the compile step folds the search index's chunk view
  into the compile state instead of relying on rows mirrored into the document
  store, which used to report pages created: 0 without an error.
• Knowledgebase.pipeline_id is a real top-level dataset column now: it is
  stored on create/update, returned in the dataset payload, and drives the
  generation gate exactly like the upstream column.
• Clearing the wiki answers the upstream confirmation copy and returns the page
  to the empty state.

v0.47.0 — Compilation templates speak the v0.27.2 vocabulary

[0.47.0] - 2026-10-20

Compilation templates speak the v0.27.2 vocabulary

The knowledge-compilation template palette and editor were re-audited against RAGFlow
v0.27.2 and brought to the current contract.

• The built-in palette is the v0.27.2 seed set: eight templates in RAGFlow's own order,
  the graph-based wiki renamed from artifacts to wiki, the mode, instruction and
  example fields included, and the retired empty seed dropped.
• Wiki blueprints (page-structure presets) carry the v0.27.2 example field; the editor's
  Blueprint selector fills the instruction/example pair from the chosen preset or the
  built-in custom defaults.
• The template editor now covers the whole upstream form: template name, kind, global
  rules, wiki entity/topic mode, RAPTOR tree settings for tree templates, re-chunk rules
  for the other kinds, per-section field cards with an add/edit field dialog, blueprint
  selection for wiki templates, a JSON preview sheet and the Back/Save footer.
• Template validation answers the upstream wording: wiki mode enum, claim/concept field
  requirements, the group scope rules and the "a wiki template cannot be combined"
  message.
• The template-group HTTP answers now use RAGFlow's envelopes: success is
  {code:0, message:"success", data:...} on HTTP 200, a missing request field is a 101
  with the upstream argument wording, and validation or not-found errors are HTTP 200
  with code 102 and no data key. The builtin palette and wiki-preset routes carry the
  same success message.
• Existing template groups saved under the old artifacts kind are migrated to wiki on
  startup, with mode=entity filled in, so no saved group is lost.

v0.46.0 — the document API answers like RAGFlow

[0.46.0] - 2026-10-12

The document API answers like RAGFlow — field for field

The document surface of the RAGFlow v0.27.2 HTTP API was replicated end to end: every
route contract was probed read-only against a live v0.27.2 instance, and each answer —
message text, numeric code, response shape — was folded back into RayRAG.

• Document listing speaks the upstream query language: keywords, page/page_size,
  orderby, desc, types and run filters, with the upstream answers for invalid
  filters and the {total, filter} facet map for type=filter. Rows carry the localised
  dates, the short run vocabulary (UNSTART/RUNNING/CANCEL/DONE/FAIL alongside
  the numeric value), the aggregated token count and the derived document type; the UI
  pages through large datasets transparently.
• The document update route (PATCH /datasets/{id}/documents/{id}) is the full field
  face — name, chunk_method, enabled, parser_config, pipeline_id, meta_fields —
  with RAGFlow's validation wordings reproduced verbatim (immutable chunk/token counters,
  the 255-byte name rule, extension and duplicate checks, automatic parser-type guards). A
  successful update answers with the complete document object.
• A rename or parser-method change resets the parse state atomically: progress cleared,
  chunk counters released, the document object reloaded.
• The metadata-config routes validate raw request bodies like upstream — an empty body
  answers metadata is required with code 101, malformed payloads answer the
  raised-exception envelopes — and unknown or foreign datasets get the permission wording.
• The ingest route, batch status, thumbnails and delete follow the probed upstream
  semantics, edge cases included; application errors ride HTTP 200 with numeric codes,
  matching RAGFlow.