Flora — Automated Content Generation

Asset generation tools for Flora: Flower of the Day — the iOS app that celebrates a new flower every day.

This repository contains two tools that do the same job — generating production-ready Xcode assets from a Latin name — but with fundamentally different architectures.

---

flora_agent — Agentic version

Gemini acts as an autonomous agent: it decides which tools to call, in what order, and how to recover from failures. The Python code provides tools; Gemini orchestrates them.

cd flora_agent
python agent.py "Rosa canina"

How it works

At each turn, Gemini sees the full conversation history (all previous tool calls and their results) and autonomously decides the next step. There is no hardcoded step sequence.

Tool	What it does
fetch_wikipedia	Fetches common name, habitat, etymology, cultural info
fetch_care_info	Scrapes PFAF for hardiness, soil and light; Gemini fallback
fetch_photos	Downloads up to 4 CC-licensed photos from Wikimedia; Gemini picks the best
process_home_image	Removes background via Gemini → transparent PNG 492×492
process_info_image	Compresses photo to <1 MB JPEG
generate_lock_image	Generates monochromatic botanical icon via Gemini
enhance_and_translate	Writes poetic copy and translates into DE, FR, ES, IT, ZH, JA
update_dataset	Upserts the flower into flowers.dataset/flowers.json

If a tool fails (e.g. background removal on the first photo), Gemini sees the error and decides to retry with the next candidate — no explicit retry logic in the code.

Setup

pip install -r ../flora_pipeline/requirements.txt

Add a .env at the repo root:

GEMINI_API_KEY=your_key_here

---

flora_pipeline — Pipeline version

A sequential Python script. Each step calls the next in a fixed order. Simpler, faster, and easier to debug.

cd flora_pipeline
python pipeline.py "Rosa canina"

Runs 7 steps end-to-end:

Step	What it does
1 — Wikipedia	Fetches common name, description, habitat, etymology, cultural info
2 — Care info	Scrapes PFAF for hardiness, soil and light requirements
3 — Photos	Fetches up to 4 CC-licensed candidates from Wikimedia; Gemini picks the best
4 — Image processing	Background removal on home.png via Gemini; compression for info.jpg
5 — Lock screen	Generates a monochromatic botanical icon via Gemini (linocut style)
6 — Content & translations	Enhances English copy and translates into DE, FR, ES, IT, ZH, JA
7 — Dataset	Upserts the flower into flowers.dataset/flowers.json

To process a batch:

python run_batch.py

Setup

pip install -r requirements.txt

Add a .env at the repo root (or inside flora_pipeline/):

GEMINI_API_KEY=your_key_here

Utilities

python clean.py   # wipe results.xcassets/ for a fresh run

---

Output structure

Both tools produce the same output layout:

results.xcassets/
  <slug>.imageset/
      home.png              ← transparent-bg photo, 492×492
  <slug>-info.imageset/
      info.jpg              ← compressed photo, <1 MB
  <slug>-lock.imageset/
      lock.png              ← Gemini-generated icon, 200×200
  flowers.dataset/
      flowers.json          ← shared array, upserted by latinName

Get your API key from Google AI Studio.

---

Pipeline vs Agent — which one to use

The pipeline makes roughly 5 Gemini calls per flower; the agent makes ~13, since each orchestration turn resends the full conversation history. For a repetitive, well-defined workflow the pipeline is cheaper and faster.

However, the agent evaluates every step on its own: it reads the result, decides whether it's good enough, and adapts — retrying a failed photo, skipping a broken step, or changing strategy without any hardcoded logic. flora_agent was built to explore that autonomy.