feat: Gemini embedding support via GEMINI_API_KEY env var (zero-config, free tier)

[flychicken067]

What this does

Adds Gemini embedding support as a zero-config free alternative to OpenAI.
Set GEMINI_API_KEY and GBrain uses Gemini automatically — no config file changes, no gbrain init flags.

Why env-var auto-detection instead of config flags

The simplest path to a working brain for users without an OpenAI key:

export GEMINI_API_KEY=your_key   # from aistudio.google.com — free
gbrain embed --stale              # just works
gbrain query "your question"      # just works

No --provider flags. No config file edits. Existing OpenAI users are unaffected.

Changes (2 files only)

src/core/embedding.ts — multi-provider with auto-detection:

• Priority: GEMINI_API_KEY → OPENAI_API_KEY → clear error with link to get a key
• gemini-embedding-001 returns 3072 dims by default; sliced to 1536 — no schema changes
• Exponential backoff (4s base, 120s cap, 5 retries) matching OpenAI path
• Free tier: 1500 req/day, sufficient for personal knowledge bases

src/core/search/hybrid.ts — fix key detection:

• Was: if (!process.env.OPENAI_API_KEY) (skipped vector search with Gemini key)
• Now: if (!process.env.OPENAI_API_KEY && !process.env.GEMINI_API_KEY)

Real-world test

Tested on a 48-page Chinese/English knowledge base:

• English query → returns Chinese article as top result via semantic similarity
• Cross-language search works because Gemini embedding space is language-agnostic
• gbrain doctor passes, 100% embedding coverage

$ gbrain query "knowledge loss when employees leave"
[0.0333] article_17 -- 字节跳动的「飞书知识库」解决不了一件事...

English question surfaces Chinese article. This is the core value of semantic over keyword search.

Relation to #58

PR #58 takes a config-based approach (--provider, outputDimensionality param).
This PR takes an env-var approach (detect key, auto-select provider).
Both maintain 1536-dim compatibility. Different UX tradeoffs — happy to consolidate if preferred.

Verification

export GEMINI_API_KEY=your_key
gbrain embed --stale
gbrain query "test question"
gbrain doctor  # should show 100% coverage

Get a free Gemini key: https://aistudio.google.com/apikey

🤖 Generated with Claude Code

[prasadus92]

+1, applied this locally on v0.12.0 and confirmed it works cleanly.

Setup on Alfred (Mac Mini dev box):

gh pr checkout 89 -R garrytan/gbrain
bun install
GEMINI_API_KEY=AIza... bun ./src/core/embedding.ts
# → dim: 1536 ✓

Why this matters for us:

• Solo-founder budget sensitivity on pay-per-token inference (OpenAI / Anthropic)
• We already have GEMINI_API_KEY wired for Hermes Agent; zero extra config to share the same free-tier key with gbrain
• 1536-dim slice preserves the existing schema — no DB migration, works with PGLite local brain out of the box

Minor observation (not blocking): the 1536 dims, free tier comment on line 5 is slightly ambiguous — the key docstring claim is that gemini-embedding-001 returns 3072 and we slice to 1536 to match the schema, which is what the code actually does (line 58). Worth a small clarification in the comment block so the "1536 dims" line reads as "we slice 3072→1536 for schema compat" rather than "the model returns 1536 natively." Not worth delaying a merge over.

Would love to see this land — there are 8+ open PRs in this area (#58, #59, #99, #134, #147, #150, #172, #178) which suggests real demand for multi-provider embeddings. The zero-config env-var approach in this PR is the cleanest UX of the bunch.

Happy to help with any reviewer asks.