[Feature]: Embedding model from url endpoint

[Whamp]

Problem Statement

Thanks for this project, i use it a lot and think it's fantastic. I host my own embedding model in my personal tailnet and would love to be able to use it as the model codegraph uses for codegraph embed . and i wanted to see if that's a feature that might be in scope or if the plan is to keep it limited to only local embed models ?

Proposed Solution

expose a new embedding-provider that is user configurable.

Alternatives Considered

No response

Area

Semantic search / embeddings

Additional Context

codegraph embed

[carlos-alm]

Thanks for the kind words, and for the well-scoped ask — self-hosting your own embedding model on your tailnet is a completely reasonable thing to want, and it's a gap on our end, not a "no."

This is prioritized. We're picking it up as a proper feature (configurable embedding provider for codegraph embed), not just noting it for later. Implementation plan in the comment below.

[carlos-alm]

Plan: configurable remote/self-hosted embedding provider

Goal: let codegraph embed call a user-configured HTTP endpoint (OpenAI-compatible /embeddings shape — covers self-hosted servers like text-embeddings-inference, llama.cpp's server, LM Studio, Ollama's OpenAI-compat endpoint, etc.) instead of only the bundled local model.

No new npm dependencies required. This is buildable entirely on what's already in the repo:

1. HTTP transport — Node's built-in global fetch (repo requires Node >=22.12, which ships fetch natively). No axios/undici/node-fetch needed.
2. Config scaffolding already exists — DEFAULTS.llm (src/infrastructure/config.ts) already declares { provider, model, baseUrl, apiKey, apiKeyCommand }, mirrored in src/types.ts. resolveSecrets() already shells out via execFileSync to resolve apiKeyCommand into apiKey — the exact pattern needed for auth against a remote endpoint. It's currently unwired to any consumer.
3. Clean seam to implement against — the local embedding generator (src/domain/search/models.ts) exposes embed(texts: string[], modelKey) -> { vectors: Float32Array[], dim: number }, a provider-agnostic interface. A remote provider is a new module implementing the same signature via a fetch POST.

Work items:

• Add baseUrl to ENV_LLM_MAP in config.ts (currently only provider/apiKey/model are mapped from env vars) so CODEGRAPH_LLM_BASE_URL works.
• Add a new provider module in src/domain/search/ (e.g. remote-provider.ts) implementing embed(texts, modelKey) -> {vectors, dim} against an OpenAI-compatible /embeddings endpoint, using baseUrl + apiKey/apiKeyCommand from config.
• Wire embeddings.llmProvider (declared in config/types but currently unused) to select between the local HF-transformers path and the new remote path at the call site in src/domain/search/generator.ts.
• Local model (@huggingface/transformers, lazy-loaded devDependency) stays the default and fallback — no change to existing behavior when no remote provider is configured.
• Tests: unit test for the remote provider against a mocked HTTP endpoint, plus a config test for baseUrl env override and apiKeyCommand resolution in this new path.

We'll open a PR against main for this.