Forge embeddings

Forge is Voxell's hosted text-embedding API. This skill covers using it through the @voxell/forge-mcp server, which exposes two tools: `embed` (text → vectors) and `list_models`. Forge does embeddings only — it does not store text, search, or do RAG for you; you keep and query the vectors yourself.

Setup

The server runs over npx and needs one env var, FORGE_API_KEY (free key, 10M tokens, no card, at https://dash.voxell.ai). Standard MCP config:

{ "mcpServers": { "forge": { "command": "npx", "args": ["-y", "@voxell/forge-mcp"],
  "env": { "FORGE_API_KEY": "your-key-here" } } } }

Choosing a model

Call list_models if unsure. Three tiers trade quality for size/cost:

Start with turbo; move up only if retrieval quality is the bottleneck. Higher dims cost more to store and compare.

Using embed

Arguments: input (string or string[]), model (default turbo), dim (optional, Matryoshka truncation), input_type ("query" or "document", default document).

Rules that matter for quality:

• Match `input_type` to the role. Embed stored content with input_type: "document" and the

user's search text with input_type: "query". Mismatching them degrades retrieval.

• Compare with cosine similarity. Vectors are L2-normalized; rank by dot product / cosine.
• Same model + same dim on both sides. Never compare vectors from different models or dims.
• Matryoshka (`dim`). Set dim below the model default (e.g. 256) for smaller, cheaper

vectors that stay re-normalized — trade a little accuracy for ~4× less storage. Use the same dim for documents and queries.

• Batch. Pass an array of texts in one call rather than one call per text.

Typical flows

• Semantic search / RAG retrieval: embed each document chunk with input_type:"document",

store the vectors; at query time embed the query with input_type:"query" and rank stored vectors by cosine similarity. Forge returns the vectors — storage and the nearest-neighbor search are yours (a vector DB, numpy, etc.).

• Similar / duplicate detection: embed two (or many) texts, compare cosine similarity; high

similarity ⇒ near-duplicate.

• Cluster / classify: embed a batch, then cluster (k-means) or train a classifier on the

vectors.

Don't

• Don't claim Forge stores data, searches, or does RAG — it returns vectors only.
• Don't call ultra "the best/top-ranked" embedding model — say "~75+ avg task score, #4 on MTEB

English, top usable model." Don't link a leaderboard.

• Don't apply the MTEB number to turbo/pro — it's the ultra/8B model.