docling-Studio

A hotfix for 0.7.1. No backend or API change.

🐛 Fixed

The preview scrolls to the focused element once, without the page collapsing (#336).

Selecting a node in the structure tree made the stacked preview flicker and land somewhere other than the element. The preview only mounts the raster of pages inside its render window, and nothing reserved the box of the others: an off-window card collapsed to its header and re-inflated once its image decoded. Every scroll target was computed against a layout changing underneath it.

Each page frame now holds the page's own aspect ratio and natural raster width, so a card keeps its height with no image mounted. On top of that stable layout:

• the focus scroll measures the frame rather than the image, so a bbox on a page far outside the window is centred on the first click, not on a later load event;
• the three watchers a single focus wakes (highlight, focus tick, current page) coalesce into one scroll per flush instead of three competing smooth scrolls;
• the page carrying the highlight keeps its raster — and its overlay — for as long as it is the target.

🛠️ Development

Worktree-friendly development stack (#324). docker-compose.dev.yml no longer binds fixed host ports or named volumes, so several worktrees can run their stack side by side: the parser is reached over the compose network, uploads and SQLite state are bind-mounted under document-parser/, and the frontend host port comes from FRONTEND_HOST_PORT (default 3000). A new .development_scripts/worktree_setup.sh prepares dependencies and copies local state into a fresh worktree; paseo.json wires it up for Paseo users.

Heads-up for existing dev setups: the dev parser no longer publishes port 8000 on the host, and its data moved from the uploads_data / db_data named volumes to document-parser/uploads and document-parser/data. Documents stored in the old volumes are not migrated. Published images and docker-compose.yml are unaffected.

Trying it

docker run -p 3000:3000 ghcr.io/scub-france/docling-studio:0.7.2-local

Full changelog: CHANGELOG.md · Compare: v0.7.1...v0.7.2

A packaging fix for 0.7.0. No feature changes.

🐛 Fixed

The published local image now contains the reasoning stack.

release.yml never passed the WITH_REASONING build argument, so docling-agent and mellea were excluded from ghcr.io/scub-france/docling-studio:0.7.0-local. On that image, POST /api/documents/{docId}/reasoning answered 503 and the Ask panel stayed hidden — 0.7.0's headline feature was unreachable from the image people actually pull. The runtime configuration panel added in #317 can toggle reasoning, but it cannot install packages.

release.yml now passes WITH_REASONING=true on the local target. The lightweight remote image is unchanged — its build stage does not declare the ARG and stays deliberately slim — and builds from source keep the opt-in default (--build-arg WITH_REASONING=true).

README reasoning section corrected. It still described the v1 graph overlay that 0.7.0 removed, and still called the stack "disabled by default". It now documents the Parse-view timeline, what the published image contains, and two upstream limits that had never been written down: docling-agent 0.6.0's chunkless loop emits only READ steps (the other kinds are reserved wire-contract values), and it reports no per-step duration, so the timeline renders in "step-order" mode by design rather than as a fallback.

⚠️ Known gap

The release gate has no reasoning coverage at all: no workflow builds a WITH_REASONING image, and release-gate.yml contains no reasoning job. A fully green gate says nothing about this feature — which is precisely how 0.7.0 shipped without it. The e2e-ui-reasoning job specified in docs/design/303-reasoning-trace-v2-parse-view.md was never created. Recorded here rather than quietly carried; worth landing before the next release that touches reasoning.

Trying it

docker run -p 3000:3000 ghcr.io/scub-france/docling-studio:0.7.1-local

For reasoning, add a reachable Ollama with your model already pulled:

docker run -p 3000:3000 \
  --add-host=host.docker.internal:host-gateway \
  -e REASONING_ENABLED=true \
  -e OLLAMA_HOST=http://host.docker.internal:11434 \
  -e REASONING_MODEL_ID=gpt-oss:20b \
  ghcr.io/scub-france/docling-studio:0.7.1-local

On native Linux Docker, --network=host with OLLAMA_HOST=http://localhost:11434 also works.

Full changelog: CHANGELOG.md · Compare: v0.7.0...v0.7.1

This release replaces the standalone reasoning page with a trace surface built into the Parse view, where every citation resolves to a real element of your PDF. It also makes the reasoning stack configurable at runtime, and considerably slims down what the app tries to be.

✨ Added

Reasoning Trace v2 — timeline + Ask panel in the Parse view (#303)

Reasoning runs inside the Parse view of an analysis (Analyses → open a run), replacing the standalone /reasoning page.

• A conversational Ask panel (composer + turn cards) drives the runs.
• Each step carries a typed kind. The wire contract defines PLAN, RETRIEVE, RERANK, READ, VERIFY, ANSWER and MAP, reserved end-to-end so future agent phases render without a contract change. docling-agent 0.6.0's chunkless RAG loop emits only READ, so today every step in a real trace is a READ.
• Timing: the run's wall-clock total is captured server-side and is real. Per-step durations are not reported by docling-agent 0.6.0, so the timeline renders in its uniform "step-order" mode with an explicit footnote rather than drawing bars that would be fiction. Per-step bars light up automatically once docling-agent#42 lands.
• The part that works today, and is the point of the release: focusing a step reveals and scrolls the linked element across the PDF preview, the document tree and the Parse view — page flip, bbox highlight, ancestor auto-expand, and reverse selection. Citations are anchored in the document, not printed beside it.

Backend trace projection (#303)

A pure domain.trace_builder projects the raw docling-agent ReasoningResult onto a debugger-facing ReasoningTrace. A new ReasoningService owns the run (query validation, latest-analysis lookup, wall-clock timing), so api/documents/{id}/reasoning no longer reaches into app.state directly.

Runtime reasoning configuration from the admin panel (#317)

A Reasoning section in /settings enables/disables reasoning and sets the Ollama URL, default model and RAG max_iterations — without a backend restart. The runner and service are rebuilt in place and /api/health follows, so the Ask surface appears and disappears live.

• Overrides persist in a new SQLite app_settings table. Env vars become bootstrap defaults; each field reports whether its value came from env or db, and reset to environment drops the overrides.
• A test connection action probes the host and turns the model field into a select of the models actually installed.
• Read-only diagnostics (docling-agent version + import path, provider, availability) make boot failures visible from the UI instead of the container logs.
• New endpoints: GET / PUT / DELETE /api/config/reasoning and POST /api/config/reasoning/test. On HuggingFace deployments only GET is served — PUT, DELETE and the connection probe return 403, the probe included because it is a server-side request to a user-supplied URL.
• max_iterations is now actually threaded into DoclingRAGAgent (new REASONING_MAX_ITERATIONS, default 5).

Analyses workspace

A dedicated Analyses section lists every run and opens its saved visualization independently of the document preview. A document can carry several analyses, and this is where the Parse view (and therefore reasoning) now lives.

🔧 Changed

• The document workspace is preview-only. /docs/:id renders the document header and page preview. Chunking and ingestion controls are gone from it, related analyses are listed separately, and deleting a document explicitly confirms the analysis cascade.
• Reasoning backend ported to the public run_with_trace() surface (#303). docling-agent is pinned to 0.6.0 and the adapter binds to a public, version-asserted API instead of the private _rag_loop(). A missing or unusable agent is detected *

0.6.2 — audit-driven release. Closes the 4 CRIT + 4 MAJ flagged by the 0.6.1 audit and the 0.6.2 release-gate audit.

Highlights

• Image size cut by ~half — latest-local drops from 6.04 GB to 3.11 GB (-48%), and a new slim latest-remote ships at 547 MB. Details below.
• docling-serve container added under the remote compose profile — the backend now ships a slim image that delegates conversion to quay.io/docling-project/docling-serve-cpu:v1.21.0 over HTTP. Zero HuggingFace Hub dependency in build or runtime.
• Backend + embedding-service migrated from pip to uv. Single source of truth is now pyproject.toml + uv.lock. BREAKING for anyone bootstrapping from requirements*.txt.
• BAKE_MODELS default flipped from true to false in both Dockerfiles. The only sanctioned HF touch in the project is now the latest-local GHCR build in release.yml. BREAKING for operators building from source — pass --build-arg BAKE_MODELS=true explicitly to restore the old behaviour. See docs/architecture/huggingface-dependency-map.md.
• Reasoning stack made opt-in via --build-arg WITH_REASONING=true (#254). BREAKING — without the flag, /api/reasoning responds 503.

Pre-built images — major size reduction

The latest-local image dropped by ~48% (-2.93 GB) versus 0.5.1, and the new remote variant ships at 547 MB — about an order of magnitude smaller than 0.5.1's only published variant. Pulls are faster, hosts use less disk, and cold starts are quicker.

Where the bytes went:

• uv migration (#254) — smaller resolved venv vs pip, deterministic lockfile, no transitive bloat that nobody noticed under pip's resolver.
• Explicit CPU-only torch via [tool.uv.sources] — drops ~3 GB of useless CUDA wheels that used to come along for the ride on latest-local.
• Reasoning made opt-in — docling-agent + mellea + their LLM SDK transitive weight move out of the default image. Rebuild with --build-arg WITH_REASONING=true to keep them.
• BAKE_MODELS opt-in for non-release builds — only release.yml builds the local target with the ~1.3 GB Docling checkpoint bake; the remote variant carries none.
• Hardened .dockerignore — leaner build context, no spurious layer churn on every rebuild.

Tags published

Both targets are tagged three ways for pinning flexibility:

• ghcr.io/scub-france/docling-studio:0.6.2-remote / 0.6-remote / latest-remote
• ghcr.io/scub-france/docling-studio:0.6.2-local / 0.6-local / latest-local

Added

• Docling Serve container wired into compose (#audit-10, dd1962e, bc9b4f8): new docling-serve service under the remote compose profile, pinned to quay.io/docling-project/docling-serve-cpu:v1.21.0. Pairs with CONVERSION_MODE=remote so the backend image stays slim and talks to a Docling Serve container over HTTP — no HF Hub dep on the build or runtime path. CI/release-gate E2E jobs now use this combo.
• HuggingFace dependency map (docs/architecture/huggingface-dependency-map.md): exhaustive list of every HF call site in the project, the single sanctioned touch point (release.yml → latest-local), and the maintenance rule that every new HF dep must be opt-in.

Changed

• Backend + embedding-service migrated from pip to uv (#254, 4d9bcf6): document-parser/requirements*.txt and embedding-service/requirements.txt removed; uv.lock plus `pyp