danvics/docling-studio
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

docs(#254): flesh out design doc + sync README image sizes
Some checks failed
Auto-close issues on release branch merge / Close referenced issues (push) Has been cancelled
Details

Browse source 
- Design doc 254: status Draft → Implemented; fill §4 (context),
  §5 (per-layer impact), §6 (alternatives), §7 (env vars / build-args),
  §8 (risks mapped to audit dims), §9 (testing + measured sizes),
  §10 (rollout), §11 (deferred follow-ups), §12 (PR + Docling tools link).
- README image-variants table: latest-local 1.9 GB → 3.2 GB (baked, with
  the BAKE_MODELS=false escape hatch documented), latest-remote 270 MB →
  585 MB (real measure on the new build).
This commit is contained in:
Pier-Jean Malandrino 2026-05-06 14:06:52 +02:00
parent 4088a8b8dd
commit 92bba34649
2 changed files with 162 additions and 112 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
README.md
View file

@ -117,8 +117,8 @@ Open [http://localhost:3000](http://localhost:3000), upload a PDF, and get resul
| Variant | Image tag | Size | Description | | Variant | Image tag | Size | Description |
|---------|-----------|------|-------------| |---------|-----------|------|-------------|
| **local** | `latest-local` | ~1.9 GB | Full — runs Docling in-process, CPU-only | | **local** | `latest-local` | ~3.2 GB | Full — runs Docling in-process, CPU-only, model checkpoints baked in (instant first run). Pass `--build-arg BAKE_MODELS=false` to drop to ~1.9 GB and let Docling download models on first request. |
| **remote** | `latest-remote` | ~270 MB | Lightweight — delegates to an external [Docling Serve](https://github.com/DS4SD/docling-serve) instance | | **remote** | `latest-remote` | ~585 MB | Lightweight — delegates to an external [Docling Serve](https://github.com/DS4SD/docling-serve) instance |
For remote mode: For remote mode:

270
docs/design/254-optim-taille-image-latest-local.md
View file

@ -30,11 +30,11 @@ the linked issue. Everything else is on the author.
- **Title on issue:** [ENHANCEMENT] Optim taille image latest-local (sortir reasoning, multi-stage, dockerignore) - **Title on issue:** [ENHANCEMENT] Optim taille image latest-local (sortir reasoning, multi-stage, dockerignore)
- **Author:** Pier-Jean Malandrino - **Author:** Pier-Jean Malandrino
- **Date:** 2026-05-06 - **Date:** 2026-05-06
- **Status:** Draft - **Status:** Implemented
- **Target milestone:** 0.6.0 — Doc-centric ingest - **Target milestone:** 0.6.0 — Doc-centric ingest
- **Impacted layers:** <backend: domain | api | services | persistence | infra> · <frontend: features/<name> | shared | app> · <e2e> · <infra/CI> - **Impacted layers:** infra/CI (Dockerfile, requirements split, compose) · docs (README, this doc)
- **Audit dimensions likely touched:** <pick from: Hexagonal Architecture · DDD · Clean Code · KISS · DRY · SOLID · Decoupling · Security · Tests · CI/Build · Documentation · Performance> - **Audit dimensions likely touched:** CI/Build · Performance · Decoupling · Documentation
- **ADR spawned?:** <no> *(write an ADR when choosing a library, moving a boundary, or deciding **not** to do something — see `docs/architecture/adr-guide.md`)* - **ADR spawned?:** no *(no load-bearing library/boundary change — Docker layout choice is locally scoped)*
--- ---
@ -82,36 +82,35 @@ for you, badly.
## 4. Context & constraints ## 4. Context & constraints
<!--
The surrounding reality the design has to live in.
### Existing code surface ### Existing code surface
List the modules / files / stores this change touches. Prefer concrete paths
over prose:
- Backend: document-parser/<layer>/<file>.py
- Frontend: frontend/src/features/<name>/{store,api,ui}.ts|.vue
- Persistence: document-parser/persistence/<repo>.py + schema in database.py
- E2E: e2e/<feature>.feature
### Hexagonal Architecture constraints (backend) - `document-parser/Dockerfile` — the multi-target file (`base` → `remote` / `local`)
The domain layer has zero imports from api / persistence / infra, and - `document-parser/requirements.txt` — shared deps for both targets
defines ports (abstract protocols) that `infra/` adapters implement. - `document-parser/requirements-local.txt` — additive Docling stack for the local target
Persistence imports only from domain. API never imports persistence - `document-parser/.dockerignore` — minimal exclusion list
directly — it goes through services. Call out any change that crosses - `docker-compose.yml` / `docker-compose.dev.yml` — both reference `target: ${CONVERSION_MODE:-local}`
these lines or adds / moves a port. - `document-parser/infra/docling_agent_reasoning.py` — already guards with `deps_present()`, so removing the deps from the standard image degrades gracefully
No domain / API / persistence / services code is touched. No SQLite migration. No frontend change. No e2e change.
### Hexagonal Architecture constraints
None crossed. The change lives entirely in `infra/CI` (Docker + requirements files). Domain/API/services/persistence are untouched and the `LocalConverter` / `ReasoningRunner` ports keep their existing shapes — only the deployment artefact changes shape, not the code.
### Deployment modes ### Deployment modes
Docling Studio ships two images (`latest-local`, `latest-remote`) driven by
`CONVERSION_ENGINE` — and a HF Space deployment on top of `latest-remote`. - `latest-local` (in-process Docling) — affected: split into "models baked" default and `BAKE_MODELS=false` slim variant.
State which modes this design supports, which it does not, and how the - `latest-remote` (delegates to Docling Serve) — affected: also slimmed (it inherited the reasoning deps via the shared `requirements.txt`).
frontend's feature flags (`chunking`, `disclaimer`) are affected. - HF Space — uses `latest-remote`, so it gets the size win for free without any HF-specific change.
- Frontend feature flags (`chunking`, `disclaimer`, `reasoningAvailable`) are unaffected; `reasoningAvailable` keeps reflecting `deps_present()` so the sidebar entry hides automatically when the standard image is used.
### Hard constraints ### Hard constraints
Compatibility (SQLite schema, API contract, Pydantic DTOs), deadlines
(milestone due date), deployment target (Docker Compose, HF Space), - **No SQLite or API contract change** — additive build-only.
performance budget (matters for Performance audit), license / privacy - **No Pydantic DTO change.**
(matters for Security audit). - **Backwards-compatible runtime behaviour** — the same `CONVERSION_ENGINE` toggle drives the same code paths; missing reasoning deps were already handled by `deps_present()`.
--> - **CI / GHCR push pipeline must keep working** — both `latest-local` and `latest-remote` tags continue to be built, with the same target names.
- **Performance budget** — image size budget per the issue is −30 %; achieved −48 % (baked) / −69 % (slim) on local and −90 % on remote (see §9 / §10).
## 5. Proposed design ## 5. Proposed design
@ -166,18 +165,64 @@ valuable than pseudocode.
### 5.1 Domain ### 5.1 Domain
Untouched.
### 5.2 Persistence ### 5.2 Persistence
Untouched.
### 5.3 Infra adapters ### 5.3 Infra adapters
Untouched at the Python level. The `LocalConverter`, `ServeConverter`, and `DoclingAgentReasoningRunner` adapters keep their current contracts. The only infra change is at the **deployment** layer:
- `requirements.txt` no longer carries `docling-agent==0.1.0` and `mellea==0.4.2`. They move to a new `requirements-reasoning.txt` (opt-in).
- `Dockerfile` is rewritten as a multi-stage build:
```
python:3.12-slim
│
┌─────────────┴─────────────┐
▼ ▼
builder-remote builder-local
│ pip install │ pip install torch torchvision (--index-url cpu)
│ -r requirements.txt │ pip install -r requirements-local.txt
│ │ if WITH_REASONING: pip install -r requirements-reasoning.txt
│ │
│ (/opt/venv) │ (/opt/venv)
└────────────┬──────────────┘
▼
runtime-base (poppler + appuser + HF_HOME, no pip, no source)
│
┌────────────┴────────────┐
▼ ▼
remote (final) local (final)
COPY venv-remote apt: libgl1 + libglib2.0-0
COPY . COPY venv-local
if BAKE_MODELS: docling-tools models download
COPY .
```
- Source (`COPY . /app`) is now COPYed only in the **final** stages — a code-only edit reuses every pip-install layer.
- Two new build-args: `WITH_REASONING` (default `false`) and `BAKE_MODELS` (default `true`). Both opt-out for `local-reasoning` / slim variants respectively.
### 5.4 Services ### 5.4 Services
Untouched.
### 5.5 API ### 5.5 API
Untouched.
### 5.6 Frontend — feature module ### 5.6 Frontend — feature module
Untouched.
### 5.7 Cross-cutting (feature flags, i18n, shared types) ### 5.7 Cross-cutting (feature flags, i18n, shared types)
- `/api/health` — no schema change. `reasoningAvailable` continues to reflect `infra/docling_agent_reasoning.deps_present()`, so it correctly reports `false` on the standard `latest-local` image and `true` on a `local-reasoning` image.
- `i18n` — no string change.
- `shared/types.ts` — no type change.
## 6. Alternatives considered ## 6. Alternatives considered
<!-- <!--
@ -193,51 +238,48 @@ If one of the alternatives represents a significant architectural fork
local decision, the ADR captures the cross-cutting one. local decision, the ADR captures the cross-cutting one.
--> -->
### Alternative A — <name> ### Alternative A — Dedicated `local-reasoning` Dockerfile target (3rd stage)
- **Summary:** - **Summary:** Add a third final target `FROM local AS local-reasoning` that runs `pip install -r requirements-reasoning.txt`. CI publishes `latest-local` and `latest-local-reasoning` separately.
- **Why not:** - **Why not:** Doubles the CI surface for very little benefit, and the duplication of intent (build-arg vs target) confuses operators. A single `--build-arg WITH_REASONING=true` is enough — operators tag the resulting image as `local-reasoning` themselves if they need the distinction.
### Alternative B — <name> ### Alternative B — Bake reasoning deps into the standard image, do nothing
- **Summary:** - **Summary:** Keep `docling-agent` + `mellea` in `requirements.txt`, accept the 5+ GB image as the cost of "everything works out of the box".
- **Why not:** - **Why not:** The standard `latest-remote` image (which delegates to Docling Serve and never reasons locally) was carrying ~5 GB of unrelated CUDA + LLM SDK weights. That alone disqualifies the do-nothing path.
### Alternative C — Bake the Docling models into a separate image and mount as a sidecar volume
- **Summary:** Build a tiny "models-only" image, mount it as a read-only volume on the backend container.
- **Why not:** Adds a deployment moving piece (multi-image orchestration) for a property — instant cold start — that a single `BAKE_MODELS=true` build-arg already gives, at the cost of +1.3 GB the operator can opt out of.
## 7. API & data contract ## 7. API & data contract
<!--
Make the wire contract explicit — this is what the frontend, e2e tests,
and any external consumer will code against.
### Endpoints ### Endpoints
| Method | Path | Request | Response | Breaking? |
|--------|------|---------|----------|-----------|
| | | | | |
Remember: No change. `/api/health` keeps the same shape; `reasoningAvailable` still derives from runtime import-checks.
- API serialization is camelCase (Pydantic `alias_generator`).
- Backend internals stay snake_case.
- `pages_json` is the documented exception — it carries raw
`dataclasses.asdict()` output (snake_case).
- Health endpoint (`/api/health`) may need new fields if this design adds
a feature flag.
### Persistence schema ### Persistence schema
```sql
-- ALTER TABLE / CREATE TABLE statements, with reasoning No change.
```
### Env vars / config ### Env vars / config
No new runtime env vars. Two new **build-args** on the `local` Dockerfile target:
| Name | Default | Allowed | Notes | | Name | Default | Allowed | Notes |
|------|---------|---------|-------| |------|---------|---------|-------|
| | | | | | `WITH_REASONING` | `false` | `true` / `false` | Bundle `docling-agent` + `mellea`. Opt in to build a `local-reasoning` image. Off keeps the standard image lean. |
| `BAKE_MODELS` | `true` | `true` / `false` | Pre-fetch Docling model checkpoints into the appuser HF cache. Off makes the image ~1.3 GB lighter at the cost of a one-time download on first conversion. |
Compose forwards `WITH_REASONING` from the host env (`WITH_REASONING=true docker compose up --build`).
### Breaking changes ### Breaking changes
Enumerate anything a consumer must change. If there are none, say so
explicitly — "additive only" is a useful commitment.
-->
## 8. Risks & mitigations **Additive only at the deployment level. Two operational expectations change** — both intentional:
1. Anyone running the standard `latest-local` image with `REASONING_ENABLED=true` will see `reasoningAvailable=false` from the API and the Reasoning sidebar entry will hide. To restore: rebuild with `--build-arg WITH_REASONING=true`. (Already documented in README.)
2. The previously-added `hf_cache` named volume in compose is removed. Models are now baked at build time; a leftover `hf_cache` volume from a prior `up` is a no-op (orphan), safe to `docker volume rm`.
<!-- <!--
One row per non-trivial risk. Map each to an audit dimension so the One row per non-trivial risk. Map each to an audit dimension so the
@ -264,81 +306,87 @@ carefully.
| Risk | Audit dimension | Likelihood | Impact | How we notice | Mitigation / rollback | | Risk | Audit dimension | Likelihood | Impact | How we notice | Mitigation / rollback |
|------|-----------------|------------|--------|---------------|------------------------| |------|-----------------|------------|--------|---------------|------------------------|
| | | | | | | | A future transitive dep silently re-pulls a CUDA torch and inflates the image again | Performance · CI/Build | Medium | High | CI image-size step regresses; `docker history` shows nvidia-* layers | Pin `torch` to a `+cpu` build in the requirements files; add a CI check that fails if `nvidia-` packages appear in the venv |
| Operators relying on R&D reasoning hit `reasoningAvailable=false` after upgrading | Documentation · Decoupling | Medium | Medium | User report or 503 from `/api/reasoning` (already gated) | README + design doc explicitly call out the `WITH_REASONING=true` path; existing `deps_present()` already degrades gracefully (no crash) |
| Baked models become stale relative to a future Docling version bump | CI/Build | Low | Low | Backend logs HF cache version mismatch; conversion still works (Docling re-downloads if needed) | Models are re-fetched at every image build; bumping `docling` in requirements-local.txt naturally produces a fresh-baked image |
| A user adds custom models at runtime and loses them on container restart | Decoupling | Low | Low | User reports lost custom models | Documented: `BAKE_MODELS=false` + add a custom volume mount on `/home/appuser/.cache/huggingface` if persistence is needed |
| Multi-stage `COPY --from=builder` increases initial build time on a cold cache | CI/Build | Low | Low | CI build duration | Cold builds are sequential by design; warm builds are dramatically faster (pip layers reused on every Python edit) — net positive |
## 9. Testing strategy ## 9. Testing strategy
<!-- ### Backend — pytest
How this design will be verified. Be specific — name files / suites.
### Backend — pytest (`document-parser/tests/`) No new tests added. The change is build-only and the existing 563-test suite is the regression net (services, persistence, API contract — none touched).
- Unit: per-layer (`tests/domain/`, `tests/persistence/`, `tests/services/`)
- Integration: services wired with real aiosqlite + real adapters
- Architecture tests (if applicable): enforce import boundaries
### Frontend — Vitest (`frontend/src/**/*.test.ts`) Validation pipeline run on the branch:
- Stores: actions / getters / mocked API
- Pure helpers (e.g. `bboxScaling.ts`-style modules): deterministic
- Components only when behavior is non-trivial; do not test markup
### E2E — Karate UI (`e2e/`) ```
- Use `data-e2e` selectors — never CSS classes (see e2e/CONVENTIONS.md) ruff check . → All checks passed
- `retry()` / `waitFor()` — never `Thread.sleep()` / `delay()` ruff format --check . → 99 files OK
- Setup via API, verify via UI, cleanup via API pytest tests/ → 563 passed, 13 skipped
- Tag appropriately: `@critical` / `@ui` / `@smoke` / `@regression` / `@e2e` ```
- **Never Playwright** — Karate is the tool here.
(2 pre-existing collection errors — `pytestarch` missing in dev venv, `_encode_picture_b64` not exported — are unrelated to this branch's diff and tracked separately.)
### Frontend — Vitest
Untouched. Not run.
### E2E — Karate UI
Not in scope. The change does not affect any user-facing behaviour.
### Manual QA ### Manual QA
Steps the reviewer can run locally (`docker-compose.dev.yml` up, scenario
to reproduce). Keep it short — if the manual list is long, automate more.
### Performance / load 1. `docker compose up --build` → confirm the backend container starts, `/api/health` returns 200, `reasoningAvailable=false`.
Required when the design claims a latency / throughput / memory property, 2. Upload a small PDF and run an analysis → first conversion completes without a multi-minute model download (validates `BAKE_MODELS=true`).
or touches the conversion hot path. 3. Rebuild with `WITH_REASONING=true docker compose up --build` and `REASONING_ENABLED=true`, with Ollama reachable → `reasoningAvailable=true`, `POST /api/documents/:id/reasoning` works.
--> 4. Rebuild with `--build-arg BAKE_MODELS=false` → image is lighter; first conversion downloads models on demand.
### Performance / load — image size measurements
Measured on arm64, cold Docker cache:
| Variant | Before | After | Δ |
|----------------------------------------|----------:|----------:|-------:|
| `latest-local` (models baked, default) | 6.09 GB | 3.19 GB | −48 % |
| `latest-local` (`BAKE_MODELS=false`) | 6.09 GB | 1.89 GB | −69 % |
| `latest-remote` | 5.85 GB | 585 MB | −90 % |
Build durations (cold cache): `after-remote` ≈ 49 s, `after-local` ≈ 1 m 46 s. Warm rebuild after a Python-only edit: sub-second (pip layers reused).
## 10. Rollout & observability ## 10. Rollout & observability
<!--
How this change gets to production safely.
### Release branch ### Release branch
Which `release/X.Y.Z` is the target? Any coordination with a parallel
release (e.g. R&D branch)? Targets `release/0.6.0`. No coordination needed with parallel branches — the change is isolated to build files.
### Feature flag / staged rollout ### Feature flag / staged rollout
Does the change hide behind a flag surfaced via `/api/health`? If so, what
flips the flag, and what is the default? HF Space deployments often need No runtime feature flag. The change is hidden behind two **build-time** flags:
`deploymentMode === 'huggingface'` gating.
- `BAKE_MODELS=true` (default) — produces the standard ~3.2 GB image.
- `WITH_REASONING=true` (opt-in) — produces a `local-reasoning` variant.
Operators can roll out by re-pulling `latest-local` from GHCR; no env flip needed.
HF Space deployments are unaffected by the `local`-side build-args (HF Space uses `latest-remote`).
### Observability ### Observability
- Logs to add / extend (structured, low-cardinality keys)
- Metrics / counters (if added — call out any new Prometheus names) - No new logs, metrics, or error modes.
- New error modes to watch for in `analysis_jobs.status = FAILED` - Image size becomes a CI signal: a follow-up issue should add a step that prints the published image size to the workflow summary, so a future regression (e.g. a new dep silently re-pulling CUDA) is visible at PR review time.
### Rollback plan ### Rollback plan
The revert that is safe to apply at any time:
- Which migration is reversible? Which is not?
- Which env var flip disables the feature without a redeploy?
- Any data cleanup needed after rollback?
Link to the existing release / ops playbooks: Pure-revert. Re-deploying the previous tag (`v0.5.x`) restores the prior image. No data migration or env flip is involved. The `hf_cache` named volume (added then removed in the same release branch) is a no-op orphan after rollback — safe to ignore or `docker volume rm hf_cache`.
- Deployment: `docs/release/*` (also surfaced via `/release:deploy`)
- Rollback: also surfaced via `/release:rollback`
- Incident: `docs/operations/*` (also surfaced via `/ops:incident`)
-->
## 11. Open questions ## 11. Open questions
<!-- Resolved during implementation. Two follow-ups deferred to dedicated issues:
Things the author explicitly does not know yet, phrased as questions the
reviewer can answer or redirect. Empty is allowed once the design is
Accepted — during Draft / In review, this section is where the honest
uncertainty lives. Resolve or delete each entry before shipping.
-->
- ... - Drop `docling-core[chunking]` extra from `requirements.txt` to push `latest-remote` from 585 MB toward the historical ~270 MB. Needs verification that the `infra/local_chunker.py` path is local-only (it appears to be, but a check is warranted).
- ... - Pin `torch` to a CPU build (`torch==X.Y.Z+cpu`) and add a CI guard that fails if `nvidia-*` packages appear in the venv — concrete safeguard against the regression mode that produced the original 5.6 GB bloat.
## 12. References ## 12. References
@ -347,12 +395,14 @@ Links to everything a future reader would want.
--> -->
- **Issue:** https://github.com/scub-france/Docling-Studio/issues/254 - **Issue:** https://github.com/scub-france/Docling-Studio/issues/254
- **Related PRs / commits:** - **Related PRs / commits:** https://github.com/scub-france/Docling-Studio/pull/255
- **ADRs:** <ADR-NNN or "none planned"> - **ADRs:** none planned
- **Project docs:** - **Project docs:**
- Architecture: `docs/architecture.md` - Architecture: `docs/architecture.md`
- Coding standards: `docs/architecture/coding-standards.md` - Coding standards: `docs/architecture/coding-standards.md`
- ADR guide / template: `docs/architecture/adr-guide.md`, `docs/architecture/adr-template.md` - ADR guide / template: `docs/architecture/adr-guide.md`, `docs/architecture/adr-template.md`
- Audit master: `docs/audit/master.md` - Audit master: `docs/audit/master.md`
- E2E conventions: `e2e/CONVENTIONS.md` - E2E conventions: `e2e/CONVENTIONS.md`
- **External:** <specs, upstream issues, dashboards, third-party docs> - **External:**
- Upstream `_rag_loop` public-API replacement: https://github.com/docling-project/docling-agent/issues/26
- Docling models tooling: `docling-tools models download` (CLI shipped by `docling`)