Gemini
Gemini is Google's native multimodal model. Uniquely, it accepts video and huge context (2M+ tokens) natively. 2025 sees Gemini 2.0/3.0.
When to Use
-
Massive Context: "Here is a 1-hour video. Find the timestamp where..."
-
Multimodal Live: Real-time voice/video interaction.
-
Google Ecosystem: Integrated with Vertex AI, Search (Grounding), and Workspace.
Core Concepts
Models
-
Pro: The best all-rounder.
-
Flash: Extremely fast and cheap. High throughput.
-
Ultra: The largest reasoning model.
Grounding
Connects the model to Google Search to provide citations and up-to-date info.
Context Initial Caching
Cache the context (e.g., a massive manual) to reduce cost/latency on subsequent queries.
Best Practices (2025)
Do:
-
Use Flash for RAG: 2.0 Flash is smart enough for most RAG & cheaper/faster.
-
Use Grounding: Eliminate hallucinations by enforcing "Google Search" grounding.
-
Upload Video: Don't transcribe video manually; Gemini watches it.
Don't:
- Don't confuse with PaLM: Gemini replaced PaLM 2 completely.
References
- Gemini API Documentation