Gemini Bible

A single-page, runnable reference for the Gemini API family. Pick a model, pick a scenario, read the code, run it against your own key, read the token bill. AI Studio and Vertex AI in parity.

---

Why this exists

Every adoption conversation looks the same. A team is moving from gemini-2.5-flash to gemini-3-flash-preview and the prompt structure shifted; or they want a tool-calling sample for Vertex but the doc page shows the AI Studio shape; or they're costing out long-context against implicit caching and the pricing page is in a different tab from the code. Three browser windows, two SDK versions, half-matching examples.

Gemini Bible collapses that loop. One surface. Browse by category and scenario; the code sample, the parameter reference, the doc link, the pricing row, and the live token meter sit on the same screen.

It is the reference the author wished existed when working with customers; it is also the reference customers can fork and take home.

---

What's in it

Categories. Text and chat, Live (realtime audio/video), image generation (Nano Banana family), video generation (Veo), embeddings.

Scenarios per category. Basic call, streaming, multi-turn chat, function calling, structured output, system instructions, multimodal input, context caching, thinking budget, safety settings, batch.

Surfaces. Every sample exists in two variants — Gemini API (AI Studio, key-based) and Vertex AI (ADC, project-scoped) — with the diff highlighted. The unified google-genai SDK keeps the bodies close; the diff is usually the client constructor and a handful of parameter names.

Languages. Python is the canonical surface. TypeScript and Java appear on scenarios where the SDK shape diverges meaningfully (live sessions, streaming back-pressure, function-calling typing).

Telemetry. Every run reports input tokens, output tokens, cached tokens, thinking tokens, latency to first token, total latency, and an estimated cost using the current public rate card.

---

Quick start

# 1. Auth — set whichever surfaces you want active
export GEMINI_API_KEY=...                                  # AI Studio
gcloud auth application-default login                      # Vertex
export GOOGLE_CLOUD_PROJECT=your-project                   # Vertex
export GOOGLE_CLOUD_LOCATION=us-central1                   # Vertex (default)

# 2. One script for everything
./app.sh start         # provision deps, background backend :8165 + frontend :5142
./app.sh status        # pid + port per side
./app.sh logs          # tail -f .run/{backend,frontend}.log
./app.sh stop          # kill by pidfile, fall back to ports
./app.sh restart

# Open http://localhost:5142

BACKEND_PORT and FRONTEND_PORT env vars override the defaults.

The backend reads keys once at startup. Nothing is persisted to disk, nothing is sent off-host. The browser talks to localhost only.

---

Layout

backend/
  app/
    main.py        FastAPI entrypoint
    auth.py        Surface detection (AI Studio key, ADC, Vertex project)
    registry.py    Sample manifest loader
    runner.py      Subprocess executor with streaming + usage capture
    pricing.py     Per-model rate card → cost estimation
  pyproject.toml

frontend/
  src/
    styles/tokens.css   OKLCH design tokens
    ui/components/      Button, Chip, Panel, Kbd, Editor
    routes/             Sample browser, sample workspace
    state/              zustand stores

samples/
  text/
    basic/
      python/ai-studio.py
      python/vertex.py
      typescript/ai-studio.ts
      manifest.json
    streaming/
    chat/
    tool-call/
    structured-output/
    context-cache/
  image/
    nano-banana/
  video/
    veo/
  live/
    realtime-audio/
  embeddings/

A sample is a directory: one or more code files plus a manifest.json declaring its category, scenario, supported surfaces, target models, doc URLs, and the run command. The registry scans the tree at startup; adding a sample is a matter of dropping in a directory.

---

Design

The interface follows the design system at Agentic-Concierge/DESIGN.md — OKLCH dark-first, Geist + Geist Mono + Fraunces, spring motion, one champagne accent, Bloomberg-density telemetry. No spinners, no sparkle icons, no purple gradients.

---

Sample inventory

ID	Category	Scenario	Surfaces	Languages
text.basic	text	basic	ai-studio · vertex	python · typescript
text.streaming	text	streaming	ai-studio · vertex	python
text.chat	text	chat	ai-studio · vertex	python
text.system-instruction	text	system-instruction	ai-studio · vertex	python
text.structured-output	text	structured-output	ai-studio · vertex	python
text.thinking	text	thinking	ai-studio · vertex	python
text.tool-call	text	function-calling	ai-studio · vertex	python
text.context-cache	text	context-cache	ai-studio · vertex	python
text.multimodal-input	text	multimodal-input	ai-studio · vertex	python
live.text-roundtrip	live	text-roundtrip	ai-studio · vertex	python
image.nano-banana	image	generation	ai-studio · vertex	python
video.veo	video	generation	ai-studio · vertex	python
embeddings.basic	embeddings	basic	ai-studio · vertex	python

Adding a sample is a matter of dropping a directory under samples/ with a manifest.json and one or more code files. The registry walks the tree at startup; the UI picks it up on the next refresh.

Telemetry

Every run emits a snapshot:

• Latency — TTFT (time to first token) for streaming and Live samples, total wall time for everything else.
• Tokens — prompt, cached, output, thinking, tool_use_prompt, total. Per-modality breakdowns (TEXT / AUDIO / IMAGE / VIDEO) when the response carries them.
• Cache hit ratio — cached / prompt. Used to validate explicit and implicit cache wins.
• Cost — USD and INR estimate. Cached tokens billed at 25% of the input rate; thinking tokens at the output rate. Refresh the rate card in backend/app/metrics.py against ai.google.dev/pricing quarterly.

The last 100 runs sit in a ring buffer; GET /api/metrics returns {summary, runs} with p50/p95 across the window.

Status

v0.1. Thirteen samples spanning text, live, image, video, and embeddings. Future work tracked in the commit log; SDK shape revisited on every Gemini family bump.