GitHub - dakshjain-1616/Qwen-Lens-Studio: Multimodal AI studio powered by Qwen3.6-35B-A3B. End-to-end web app exposing visual reasoning, image captioning, and document understanding tools from a single model with side-by-side output across versions.

A multimodal AI studio built around a single Qwen vision-language model, exposed through five focused tools plus a batch runner and a persistent session log. Ship a screenshot → get code. Ship an invoice → get structured JSON. Ship two images → get a structured diff.

What's in the box

Five vision tools, all streaming:

Plus:

• Batch runner — drop many images, pick a tool, drain a 3-worker queue, export results as CSV or JSON (Document IQ CSV auto-flattens JSON keys into columns).
• Session history — every run is stored (localStorage, up to 100 per tool), searchable, restorable, exportable.
• Prompt template library — save reusable system prompts per tool; sent as a one-call override.
• Export menu on every result — Markdown, JSON, or PDF (via html2pdf.js).
• In-UI API key — paste your OpenRouter key into Settings; stored in localStorage, attached as X-OpenRouter-Key on each request. No .env edit required for non-developers.
• Monaco read-only code viewer for Code Lens output and the Document IQ Raw tab.
• Live preview button on Code Lens results opens the generated UI in a new browser tab.
• Markdown rendering for all natural-language answers (bullets, headings, tables, code fences).

Architecture

React + Vite + TypeScript SPA  ──►  FastAPI backend  ──►  Qwen via OpenRouter / Ollama / llama.cpp
       :5173 (dev)                       :8001                    (pluggable)
                   proxies /api/*

• Backend: FastAPI, app.py + 5 processor classes in processors/. One QwenClient in qwen_client.py handles streaming, thinking-trace parsing, and per-call API key / model / backend overrides.
• Frontend: React 18 + Vite + TypeScript. State: local useState per tool + one Zustand store for history and UI flags. Styling: Tailwind, dark glass theme. Markdown rendered via react-markdown + remark-gfm. Code rendered via @monaco-editor/react (read-only).
• SSE: Each tool endpoint returns text/event-stream. Per-chunk JSON: {content, thinking, is_thinking, is_complete, error}. The client distinguishes thinking from answer deltas and drives two independent panes.
• No database. History is browser-side localStorage (qls.history.v1, quota-aware with oldest-drop fallback). Config lives in qls.config.v1. Prompts in qls.prompts.v1.

Getting started

Prerequisites

• Python 3.10+
• Node 18+
• An OpenRouter API key (or Ollama / llama.cpp running locally)

Backend

cd qwen_lens_studio_1025
python3 -m venv .venv && source .venv/bin/activate   # optional
pip install -r requirements.txt

cp .env.example .env
# edit .env: OPENROUTER_API_KEY=...; QWEN_MODEL=qwen/qwen3.6-plus; MOCK_MODE=false

uvicorn app:app --host 0.0.0.0 --port 8001 --reload

For demos without an API key, set MOCK_MODE=true and the backend will stream a canned response.

Frontend

cd frontend
npm install
npm run dev -- --host 0.0.0.0
# open http://localhost:5173/

The Vite proxy forwards /api/* to http://127.0.0.1:8001 (see vite.config.ts).

Production build

cd frontend && npm run build         # emits frontend/dist
# then serve dist/ behind any static host, and run uvicorn behind that

Configuration

Per-install (.env)

Per-user (in-UI)

Click the gear icon in the top-right. Set:

• OpenRouter key (masked, stored in localStorage)
• Model ID (leave blank to use the server default from .env)
• Backend (openrouter / ollama / llamacpp)

Attached per-request as X-OpenRouter-Key, X-Model-Id, X-Backend headers; these override the server defaults for that request only.

Per-run (prompt templates)

Each tool page has an "Edit system prompt" expandable. Save named templates in the Prompt Library modal (qls.prompts.v1). Sent to the backend as the optional custom_system_prompt form field — the processor uses it in place of its default prompt when present.

Endpoints

Every tool endpoint takes multipart form-data and returns SSE.

Optional headers on every tool endpoint: X-OpenRouter-Key, X-Model-Id, X-Backend.

Project layout

qwen_lens_studio_1025/
├── app.py                        # FastAPI routes + SSE helpers
├── qwen_client.py                # QwenClient (stream + non-stream)
├── processors/
│   ├── reasoning.py
│   ├── multilingual.py
│   ├── document_iq.py
│   ├── code_lens.py
│   └── dual_compare.py
├── tests/                        # pytest suite
├── .env.example
├── requirements.txt
└── frontend/
    ├── src/
    │   ├── App.tsx
    │   ├── index.css             # Tailwind + custom utility classes
    │   ├── pages/                # Reasoning / Multilingual / Document / Code / Dual / Batch / Home
    │   ├── components/           # Layout, Sidebar, HistoryPanel, SettingsModal,
    │   │                         # PromptLibraryModal, ExportMenu, ReadOnlyEditor,
    │   │                         # CodePreview, JsonTree, DiffCompare, Markdown,
    │   │                         # ImageDropzone, MultiImageDropzone, ...
    │   ├── lib/
    │   │   ├── api.ts            # SSE client, streamForm, compressImage, makeThumbnail
    │   │   ├── useToolRun.ts     # streaming state machine
    │   │   ├── history.ts        # localStorage store, search, export, cap=100/tool
    │   │   ├── config.ts         # API key + model + backend localStorage
    │   │   ├── prompts.ts        # prompt template localStorage CRUD
    │   │   ├── export.ts         # Markdown / JSON / PDF / CSV helpers
    │   │   └── types.ts
    │   └── store/useAppStore.ts  # zustand (history, UI flags, restore queue)
    ├── package.json
    ├── vite.config.ts            # /api proxy → 127.0.0.1:8001
    └── tailwind.config.js

Testing

# backend
python3 -m pytest tests/

# frontend typecheck + build
cd frontend && npm run build

Notes

• The frontend's localStorage keys (qls.history.v1, qls.config.v1, qls.prompts.v1) are versioned — bump the .v1 suffix when introducing a breaking schema change.
• MOCK_MODE does not invoke any model; it only streams canned text and is useful for offline demos or CI.
• API keys pasted into the Settings modal live in browser localStorage and are visible to any script running on the page origin. Don't paste keys on a host you don't trust.
• For large batch runs, keep the browser tab in focus — some browsers throttle fetches in background tabs.

License

Internal / unlicensed unless otherwise noted.