Embedding & Model Providers

Each team brings its own AI endpoints. Two provider types are configured under Settings → Integration:

• Embedding Providers power semantic search across prompts, artifacts, blueprints, and memories.
• Model Providers register OpenAI-compatible LLM endpoints for the team.

Both are team-scoped, managed entirely in the app, and store API keys encrypted. Nothing is configured through server environment variables.

Owner or Admin, and a publicly routable URL

Adding or editing a provider needs the Owner or Admin role (the team.settings.update permission; see Team roles and permissions). The base URL must be publicly routable: VibeXP rejects loopback, private, and link-local addresses (including cloud metadata endpoints like 169.254.169.254) with destination_not_allowed, and the check applies to the stored provider, not just the initial connectivity test. Local development against localhost is exempt.

Self-hosting the endpoint on your own private network (Ollama, TEI, vLLM on a Docker subnet)? An operator can declare that range in OUTBOUND_ALLOWED_CIDRS so the guard allows it. See Self-Hosting.

Embedding providers

An embedding provider is any OpenAI-compatible /v1/embeddings endpoint: OpenAI, Ollama, LocalAI, vLLM, HuggingFace TEI, and similar.

Adding a provider

1. Open Settings → Integration → Embedding Providers
2. Click to add a provider and fill in:

Field	Default	Purpose
Base URL, API key, provider type	(required)	The endpoint to call. The key is stored encrypted.
Model	(required)	The embedding model id. Tagged on every vector and used as the search filter.
Chunk size	1000	How much text goes into each embedding chunk.
Chunk overlap	200	Overlap between chunks (must be smaller than the chunk size).
Concurrency	1	Max concurrent embedding requests sent to this provider.
Query / document prefix	(empty)	Optional instruction prefixes for asymmetric models (mxbai, BGE, E5).

3. Save. The provider is validated on save: VibeXP calls the endpoint and requires 1024-dimension vectors. The vector width is fixed and not configurable, so pick a model that outputs (or can be asked for) 1024 dimensions.

Once saved, embedding starts automatically for the team’s existing content.

Coverage, reprocess, and clear

The Embedding Providers page shows coverage cards: how much of the team’s content is embedded, pending, or failed. Two actions are available:

• Reprocess pending: re-enqueue anything not yet embedded.
• Clear all embeddings: delete every stored vector for the team. Content is untouched; embeddings are rebuilt on the next processing pass.

Changing a provider re-embeds everything

Changing a provider’s identity (endpoint or model) wipes that team’s vectors and re-embeds all content, so the search index stays consistent. Expect a re-embedding pass after such a change. Editing only chunk sizing, concurrency, or prefixes fills gaps instead of wiping.

Without a provider

Everything still works: content saves normally and search falls back to keyword (full-text) mode, which includes typo tolerance. Only semantic search is unavailable.

Model providers

Model providers let a team bring its own OpenAI-compatible LLM endpoint.

1. Open Settings → Integration → Model Providers
2. Add the endpoint URL, API key, and model details
3. Save. The provider’s connectivity is validated on save.

API keys are stored encrypted and can be updated or removed anytime.

AI Summary

With a model provider in place, search can summarize its top results with it. The settings are a card on the same Model Providers page, below the providers. Until the team has at least one model provider the card only says that AI Summary needs one.

Field	What it does	Allowed
Enable AI Summary	Summarize the top search results with this team’s model.	on / off
Provider	Which model provider writes the summary: Team default or a specific one.	any of the team’s model providers
Style	How detailed the summary is.	Concise, Balanced, Detailed
Results to read	How many top-ranked results the summary reads.	1–10
Response length (tokens)	Upper bound on how much the summary may write.	1–32768

Changing them needs the Owner or Admin role, like the providers themselves.

A team either has its own complete AI summary profile or none at all:

• Using instance defaults: the team has saved nothing, so it uses the defaults its operator set, with Team default as the provider, and follows along when the operator changes them.
• Customized for this team: the team’s saved values apply, and later changes to the instance defaults no longer affect it.

Reset to defaults drops the team’s profile so it inherits again. It returns to whatever the instance defaults are at that moment.

The limits of 10 results and 32768 tokens are fixed. Since v0.15.0 they are the same for every team and no instance setting narrows them. An operator still sets how much text of each result a summary reads and how long the server waits for the model, and those apply to every team: see Search & AI Summary Settings.

API access

Both provider types have team-scoped REST endpoints:

# Embedding providers
GET|POST   /api/v1/{team_id}/settings/embedding-providers
GET|PUT|DELETE /api/v1/{team_id}/settings/embedding-providers/{id}
POST       /api/v1/{team_id}/settings/embedding-providers/validate
GET        /api/v1/{team_id}/settings/embedding-providers/coverage
POST       /api/v1/{team_id}/settings/embedding-providers/{id}/reprocess
DELETE     /api/v1/{team_id}/settings/embedding-providers/embeddings

# Model providers
GET|POST   /api/v1/{team_id}/settings/model-providers
GET|PUT|DELETE /api/v1/{team_id}/settings/model-providers/{id}
POST       /api/v1/{team_id}/settings/model-providers/validate

# AI summary settings
GET|PUT|DELETE /api/v1/{team_id}/settings/ai-summary

In the AI summary settings response, max_top_n and max_output_tokens_ceiling are deprecated: since v0.15.0 they always report the fixed limits, 10 and 32768.

See API Keys for authentication.

Related

• Memory and Artifacts explain how semantic search is used day to day.
• Self-hosters: see Self-Hosting → Search and embeddings.