Nipun/gemini-image-puzzle

🧩 Gemini Image Puzzle

A password-protected FastAPI app that answers vision + text questions with Gemini Flash. Upload an image or paste an image URL, ask a question, and get an answer. Ships with built-in sample puzzles (shape-algebra, similar-triangles geometry, logic, counting, chart reading).

Two answer modes:

• —Solver (ReAct) — Gemini reasons step by step and uses its code-execution tool to write & run Python (sympy), so arithmetic and algebra are computed and verified, not guessed. Each answer is driven by a skill (Auto, Geometry, Shape Algebra, Logic, Counting, Chart, General) that injects a focused system prompt + worked example. The reasoning + code is shown in a collapsible trace. Math renders with KaTeX.
• —Quick — a single fast vision answer with no tools.

The frontend is a static HTML/CSS/JS single page styled with the nipunbatra.github.io design system (Playfair Display headings, Work Sans body, warm-red #c44536 accent, light/dark theme). FastAPI serves the page and the /api/* endpoints; this avoids the issues a long-running Streamlit app hits on Spaces.

Secrets (Space settings → Variables and secrets)

Endpoints

• —GET / — frontend (injects an access-code overlay when ACCESS_CODE is set)
• —POST /api/auth — exchange access code for a session token
• —GET /api/config — available models + puzzle catalogue
• —POST /api/ask — {question, model, image_b64 | image_url | puzzle_id} → answer
• —GET /api/puzzle/{id}/answer — intended answer for a sample puzzle
• —GET /api/health — health/session check

Local run

pip install -r requirements.txt
export GEMINI_API_KEY=... ACCESS_CODE=...
uvicorn main:app --host 0.0.0.0 --port 7860

The puzzle images in static/puzzles/ are generated by python make_puzzles.py.