# VLM Playground (PreviewSpace) — Product Requirements Document

## Summary

An internal Gradio Blocks app for rapid, structured experimentation with a Vision-Language Model (initially `dots.ocr`). It mirrors the reference playground but is deliberately minimal: stateless by default, no run history, focused on feeling model performance. Supports PDF/image upload, page preview and navigation, page-range parsing, and result views (Markdown Render, Markdown Raw Text, Current Page JSON) with preserved scroll position. Designed to run locally or on Hugging Face Spaces.

## Goals

- **Fast iteration**: Upload, prompt, parse, iterate in seconds with minimal ceremony.
- **Model-light**: Start with one model (`dots.ocr`), optional model selector later. No provider switching UI.
- **Structured output**: First-class JSON output and markdown preview.
- **Stateless by default**: No run history or persistence beyond the current browser session unless explicitly downloading.
- **Document-centric UX**: Multi-page PDF preview, page navigation, per-page execution, and page-range parsing.

## Non-Goals

- Not a full labeling platform or production extraction pipeline.
- Not a dataset hosting service or long-term data store for PHI.
- Not a fine-tuning/training product; inference playground only.
- No bounding-box drawing or manual annotation tools in v1.

## Primary Users / Personas

- **Applied Researcher / Data Scientist**: Tries different prompts/models, collects structured outputs.
- **ML Engineer**: Prototypes pipelines, compares providers, validates latency/cost.
- **Domain Expert (e.g., Clinical Analyst)**: Uses curated templates to extract specific fields.

## Key User Stories

- As a user, I can upload a PDF or image, select a template prompt, and click Parse to see Markdown and JSON results.
- As a user, I can preview pages, specify a page range to parse, and run per-page extraction.
- As a user, I can jump to a specific page index in a PDF and use Prev/Next controls.
- As a user, I can switch between result tabs (Markdown Render, Markdown Raw Text, Current Page JSON) without losing scroll position.
- As a user, I can download the results for my current session as a ZIP or JSON/Markdown.
- As a user, I can tweak the prompt and basic model settings and quickly re-run.

## UX Requirements (inspired by dots.ocr playground)

- **Left Panel — Upload & Select**
  - Drag-and-drop or file picker for PNG/JPG/PDF; show file name and size.
  - Optional Examples dropdown (curated sample docs and pre-baked prompts).
  - File ingestion for PDFs extracts page thumbnails and page count.
- **Left Panel — Prompt & Actions**
  - Prompt Template select; Current Prompt editor (multiline with variable chips).
  - Actions: Parse (primary), Clear (secondary).
  - Show prompt variables, e.g., `bbox`, `category`, `page_number`.
- **Left Panel — Advanced Configuration**
  - Preprocessing toggle (fitz-like DPI upsample for low-res images).
  - Minimal server/model config: Host/Port for local inference or a dropdown for on-host models.
  - Page selection: single page, page range, or all.
- **Center — File Preview**
  - Large page preview with pan/zoom; page navigator (Prev/Next and page picker).
  - Page jump field to go directly to page N.
- **Right Panel — Result Display**
  - Tabs: Markdown Render Preview, Markdown Raw Text, Current Page JSON.
  - Preserve scroll position when switching tabs.
  - Copy-to-clipboard and a Download Results button.

## Functional Requirements

- **File Handling**
  - Accept PDF (up to 300 pages) and images (PNG/JPG/WebP). Max upload 50 MB (configurable).
  - Extract page images for preview; store temp files locally (ephemeral) with TTL.
  - Provide page-level selection and batching.
- **Prompting**
  - Template library with variables and descriptions. Variables can be sourced from UI state (page, bbox list) or user input.
  - System prompt + user prompt fields; allow few-shot examples.
  - Presets for common tasks (layout extraction, table extraction, key-value extraction, captioning).
- **Model Support**
  - Start with `dots.ocr` via the official parser or REST endpoint.
  - Optional: dropdown to switch among `dots.ocr` model variants if present on the host. No cross-provider switching UI.
- **Execution**
  - Run per-page or whole-document, controlled by UI. Concurrency limit (default 3).
  - Timeouts and retries surfaced to UI; cancellation supported.
  - Caching: request hash on (file checksum, page, prompt, params, model) to avoid recomputation.
- **Outputs**
  - Markdown Render, Raw Markdown, and Current Page JSON.
  - Export: Download button to export combined Markdown, per-page JSONL, and all artifacts as a ZIP.
- **Examples Gallery**
  - Preloaded example docs and templates to demonstrate patterns (OCR table, K/V extraction, figure captioning, layout detection).
- **Observability**
  - Show basic runtime info (latency, model id) inline; no history or centralized logs in v1.

## Data Model (high-level)

- In-memory, per-session structures only; no database.
- **Document**: id, name, type, checksum, page count, temp storage path, created_at.
- **Page**: document_id, page_index, image_path, width, height, preview_thumbnail.
- **Template**: id, name, description, model_defaults, prompt_text, output_schema (optional JSON Schema), variables.

## JSON Output Guidance

- For structured tasks, templates may specify an output schema. The UI validates model JSON and highlights issues.
- All results stored as JSON lines per page with summary aggregation.

## Security & Compliance

- Internal-only; access requires SSO or VPN.
- Sensitive documents (e.g., PHI) processed only against approved providers/endpoints. Warn when a provider is external.
- Ephemeral storage with TTL auto-clean; configurable retention. Redact logs where needed.

## Performance Targets

- Cold start to first parse: < 10s on typical PDFs (<= 20 pages) with network providers.
- Per-page preview render: < 500ms after page image generation.
- Concurrency: default 3 parallel page requests; configurable up to 10.
- Throughput: 1,000 pages/day per user on average use without manual scaling.

## Error States & Edge Cases

- Unsupported file types or oversize files; clear messaging and guardrails.
- Pages with extreme aspect ratios or very small text; suggest preprocessing.
- Provider rate limits; exponential backoff and UI feedback.
- Invalid model JSON; surface diffs and attempt best-effort JSON repair (opt-in).

## Architecture (proposed)

- **App**: Single Gradio Blocks app (Python). No separate backend required.
- **Execution**: Use `uv run` locally. Designed to run as-is on Hugging Face Spaces.
- **Model**: `dots.ocr` via local parser or REST endpoint; configurable host/port.
- **Storage**: Ephemeral `/tmp/previewspace/*`; cleared at session end or TTL.
- **Caching**: Optional on-disk cache keyed by content hash + prompt + params + model.

## API Surface (v1)

- Pure Gradio callbacks; no public REST API. Optional: expose simple `/healthz`.

## Templates (initial set)

- **Layout Extraction**: Return list of elements with `bbox`, `category`, and `text` within bbox.
- **Table Extraction**: Return rows/columns as structured JSON; include confidence and cell bboxes.
- **Key-Value Extraction**: Extract specified fields with locations and normalized values.
- **Captioning/Description**: Summarize or caption selected regions or whole pages.

## Privacy-by-Design Defaults

- Local processing preferred where possible; clear visual indicator when sending to external APIs.
- Redaction utilities for logs; toggle to disable request logging entirely.

## Success Metrics

- Time-to-first-result after upload.
- Number of saved runs and templates re-used.
- Reduction in manual extraction time for a representative task.
- User satisfaction (quick pulse after saved runs).

## Release Plan

- **M1 (v0.1) — Core Playground**
  - Upload PDF/image; page preview and navigation.
  - Parse with one provider; show Markdown and JSON; save runs; export JSON.
  - Basic provider config (host/port/api key) and preprocessing toggle.
  - Acceptance: A user can replicate a layout extraction example end-to-end in < 2 minutes.
- **M2 (v0.2) — Templates, Regions, and Examples**
  - Template library + editor; draw/save bboxes; per-page runs; examples gallery.
  - Multiple providers; concurrency and caching; logs and token usage.
  - Acceptance: A user can create a new template with variables and run it across 10 pages with regions in one click.
- **M3 (v0.3) — Projects and Evals**
  - Projects grouping; batch runs over documents; dataset export; simple eval harness with spot checks.
  - Acceptance: A user can run a project over 100 pages and export an evaluation-ready JSONL in < 10 minutes.

## Open Questions

- Do we require strict JSON schema validation with auto-repair, or soft validation with warnings?
- What are the approved external providers for sensitive documents?
- Should we include table renderers in the UI, or keep to JSON/Markdown only?
- How long should run artifacts persist by default (e.g., 7 days)?

## Risks & Mitigations

- **External API variability**: Abstract through connectors; provide stubs/mocks for local dev.
- **Document diversity**: Offer preprocessing toggles and template variables; maintain an examples gallery.
- **Cost visibility**: Track token usage and estimated cost per run; warn when large batches are selected.

## Appendices

### Example: Layout Extraction Prompt (concept)

```text
System: You are a vision-language model that outputs structured JSON only.
User: Please output the layout information from the PDF page image. For each element, return:
- bbox: [x1, y1, x2, y2] in image pixels
- category: string label from {"title","header","paragraph","table","figure","footnote"}
- text: content within bbox
Return JSON: {"elements": [{"bbox": [..], "category": "..", "text": ".."}], "page": <number>}.
```