wehub-resource-sync
426e9eeabd
4.2 KiB
@elizaos/plugin-embeddings
A provider-agnostic ("bring your own") TEXT_EMBEDDING provider for elizaOS agents. Point one set of EMBEDDING_* variables at any OpenAI-compatible /embeddings endpoint and get embeddings independently of your chat provider.
Why
Embeddings power memory, recall, and semantic search — but they don't have to come from the same provider as the chat brain. If your agent runs on a provider that serves no good embeddings (e.g. Claude, which has no embeddings API, or Cerebras, which serves none), this plugin lets you keep your chat brain where it is and route embeddings to something that does it well:
- a personal OpenAI key (
text-embedding-3-small/-large) - Eliza Cloud embeddings
- Voyage AI (via an OpenAI-compatible proxy)
- a local TEI, Infinity, vLLM, or LM Studio server
Purely additive
The plugin activates only when you set EMBEDDING_BASE_URL or EMBEDDING_API_KEY. With neither set it never loads, so dropping it into an existing deployment changes nothing until you opt in.
What it registers
Only the embedding slots — nothing else (no text/image/audio, no actions/providers/services):
| Slot | Behavior |
|---|---|
TEXT_EMBEDDING |
Embed one text → one vector. |
TEXT_EMBEDDING_BATCH |
Embed many texts in one request → one vector each. |
Both use raw fetch (no @ai-sdk dependency) to POST to ${EMBEDDING_BASE_URL}/embeddings.
Priority
Registered at priority: 1:
local-inference @ 0 < plugin-embeddings @ 1 < Eliza Cloud @ 50
A bring-your-own endpoint beats a bare local embedder but yields to a paired Eliza Cloud. Override per-slot via runtime routing preferences if you want a different order.
Fail loudly, never fabricate
On any HTTP / config / response-shape error the handler throws — it never returns a zero or garbage vector that would silently corrupt the embedding store. The only synthetic return is the boot dimension-probe (null input), where a correctly-sized marker vector is the expected, legitimate response.
Configuration
All variables are read via runtime.getSetting(key) first, then process.env, so they are per-character overridable. They do not fall back to any chat provider's settings.
| Variable | Default | Description |
|---|---|---|
EMBEDDING_BASE_URL |
(none) | OpenAI-compatible /embeddings base URL. Required for real embedding calls — no default endpoint. |
EMBEDDING_API_KEY |
(none) | Bearer token. Omit for local servers that need no auth. |
EMBEDDING_MODEL |
text-embedding-3-small |
Model id sent as the request model field. |
EMBEDDING_DIMENSIONS |
1536 |
Vector width (see below). Sent as the request dimensions field when explicitly set. |
EMBEDDING_BROWSER_URL |
(none) | Browser-only server-side proxy URL. In a browser build the Authorization header is sent only when this is set, keeping the key off the client. |
Setting either EMBEDDING_BASE_URL or EMBEDDING_API_KEY activates the plugin.
Supported dimensions
EMBEDDING_DIMENSIONS must be one of:
384, 512, 768, 1024, 1536, 2048, 3072
Any other value throws.
⚠️ Keep the dimension stable per database
The embedding dimension is baked into your database's vector schema. Changing EMBEDDING_DIMENSIONS (or the model's native width) invalidates old-width vectors until the active database adapter reclaims them and re-embeds those memories at the active width. SQL-backed agents run that cleanup at boot through clearEmbeddingsOutsideActiveDimension(); custom stores need an equivalent path.
Example .env
Personal OpenAI key for embeddings, while chat stays on another provider:
EMBEDDING_BASE_URL=https://api.openai.com/v1
EMBEDDING_API_KEY=sk-...
EMBEDDING_MODEL=text-embedding-3-small
EMBEDDING_DIMENSIONS=1536
Local TEI / Infinity / vLLM server (no auth):
EMBEDDING_BASE_URL=http://localhost:8080/v1
EMBEDDING_MODEL=BAAI/bge-small-en-v1.5
EMBEDDING_DIMENSIONS=384
Installation
The plugin is picked up automatically when EMBEDDING_BASE_URL or EMBEDDING_API_KEY is present. To reference it explicitly in a character file:
{
"plugins": ["@elizaos/plugin-embeddings"]
}