Summary

Add Google Gemini as a first-class LLM provider in zeph-llm, supporting chat, streaming, tool use, vision, and embeddings.

Motivation

Google Gemini offers competitive models (2.5-pro, 2.5-flash, 2.0-flash) with 1M token context windows, native function calling, vision, and thinking capabilities. Adding Gemini as a provider expands Zeph's multi-provider coverage and enables orchestrator/router configurations that include Google's model family.

API Reference

• Chat: https://ai.google.dev/gemini-api/docs/text-generation
• Embeddings: https://ai.google.dev/gemini-api/docs/embeddings
• Function calling: https://ai.google.dev/gemini-api/docs/function-calling
• Models: https://ai.google.dev/gemini-api/docs/models

Key API Characteristics

• Base URL: https://generativelanguage.googleapis.com/v1beta
• Auth: API key via ?key= query parameter
• Message format: contents[].parts[] (not OpenAI-compatible)
• System prompt: separate systemInstruction field
• Role names: user / model (not assistant)
• Streaming: ?alt=sse query param (SSE)
• Tool format: functionDeclarations / functionCall / functionResponse
• Embeddings: embedContent endpoint with task type parameter

Implementation Phases

• Phase 1: Basic chat (generateContent) -- feat(gemini): Phase 1 -- basic chat (generateContent) #1593
• Phase 2: Streaming support (streamGenerateContent + SSE) -- feat(gemini): Phase 2 -- streaming support (SSE) #1594
• Phase 3: Tool use / function calling -- feat(gemini): Phase 3 -- tool use / function calling #1595
• Phase 4: Vision / multimodal (inlineData parts) -- feat(gemini): Phase 4 -- vision / multimodal (inlineData) #1596
• Phase 5: Embeddings (embedContent) -- feat(gemini): Phase 5 -- embeddings (embedContent) #1597
• Phase 6: List models + remote model discovery -- feat(gemini): Phase 6 -- list models + remote discovery #1598

Config Format

[llm]
provider = "gemini"

[llm.gemini]
model = "gemini-2.5-flash"
max_output_tokens = 8192
embedding_model = "text-embedding-004"

Vault key: ZEPH_GEMINI_API_KEY

Detailed Plan

See .local/plan/gemini-provider.md