feat: add OpenAI-compatible and Gemini API embedding support

[ysys143]

What

Adds external embedding API support as an alternative to local GGUF models. Three environment variables activate API mode:

Variable	Description	Example
`QMD_EMBED_API_URL`	Base URL (presence activates API mode)	`https://api.openai.com/v1\`
`QMD_EMBED_API_KEY`	API key	`sk-...`
`QMD_EMBED_API_MODEL`	Model name	`text-embedding-3-small`

API type is auto-detected from the URL — no extra config needed:

• URL contains `googleapis.com` → Gemini (`batchEmbedContents` endpoint)
• Otherwise → OpenAI-compatible format (works with Ollama, LM Studio, OpenAI, etc.)

Usage

```bash

OpenAI

export QMD_EMBED_API_URL="https://api.openai.com/v1"
export QMD_EMBED_API_KEY="sk-..."
export QMD_EMBED_API_MODEL="text-embedding-3-small"
qmd embed && qmd vsearch "test query"

Gemini (free tier)

export QMD_EMBED_API_URL="https://generativelanguage.googleapis.com/v1beta"
export QMD_EMBED_API_KEY="AIza..."
export QMD_EMBED_API_MODEL="gemini-embedding-001"
qmd embed && qmd vsearch "test query"

Ollama (OpenAI-compatible)

export QMD_EMBED_API_URL="http://localhost:11434/v1"
export QMD_EMBED_API_MODEL="nomic-embed-text"
qmd embed && qmd vsearch "test query"
```

Why

Local GGUF embedding is great for privacy and offline use, but has tradeoffs:

• Requires downloading a model (~300MB+)
• Slower on CPU, needs Metal/CUDA for reasonable speed
• Limited model selection

API mode is useful when:

• Speed matters (cloud APIs are 2–5× faster in practice)
• You want higher-dimensional embeddings (3072 vs 768)
• Running in CI or on a machine without GPU

Implementation notes

• `formatQueryForEmbedding` / `formatDocForEmbedding` return raw text when `modelUri` starts with `api:` — cloud models don't use nomic-style task prefixes
• Model stored in `content_vectors.model` as `api:` — switching between local and API requires re-embedding (same as switching local models, existing behavior)
• Vector dimension is set dynamically on first embed, so any dimension works without schema changes

Benchmark (63 chunks, Korean+English, M3 Max)

Model	Per chunk	Cost (63 chunks)	Dims
embeddinggemma-300M local	72ms	$0	768
text-embedding-3-small	16ms	$0.000290	1536
text-embedding-3-large	13ms	$0.001882	3072
gemini-embedding-001	38ms	free	3072
gemini-embedding-2-preview	41ms	free	3072

Full benchmark results + script in `docs/embedding-benchmark.md` and `scripts/benchmark-embed.ts`.

Relation to other PRs

#406 and #415 also address remote embedding — worth noting the differences:

• feat: add remote embedding provider support (Gemini, OpenAI) #406 introduces a new `src/remote-embedding.ts` class and `index.yml` config file for provider setup
• Add external model provider support for embedding & reranking #415 adds an OpenAI-compatible remote server mode (focused on oMLX/Apple Silicon)
• This PR uses env vars only (`QMD_EMBED_API_URL` / `QMD_EMBED_API_KEY` / `QMD_EMBED_API_MODEL`), consistent with QMD's existing configuration pattern (`QMD_EMBED_MODEL`, `QMD_EMBED_CTX_SIZE`, etc.), and integrates directly into `llm.ts` with a minimal diff (~130 lines added). No new files or config formats needed.

Happy to collaborate with the other authors or adjust the approach based on feedback.