llama-index-embeddings-forge 0.1.0

LlamaIndex embeddings for Forge — Voxell's text-embedding API (turbo/pro/ultra; ultra = Qwen3-Embedding-8B, ~75+ avg MTEB, #4 English).

llama-index-embeddings-forge

LlamaIndex embeddings for Forge — Voxell's hosted text-embedding API.

Why Forge

One API, three tiers — pick your point on the quality/cost curve:

Model	Dim	Notes
turbo	1024	fast, low cost
pro	2560
ultra	4096	Qwen3-Embedding-8B; ~75+ avg task score on MTEB, currently #4 on MTEB (English) — the top usable model (the three above are research-only)

Matryoshka (MRL) dimensions are real: truncated vectors are re-normalized, so a shorter dim is a unit-norm prefix of the full vector — smaller index, minimal quality loss. Forge logs request metadata only (model, tokens, latency) — never your text or vectors.

Install

pip install llama-index-embeddings-forge

Usage

from llama_index.embeddings.forge import ForgeEmbedding

# FORGE_API_KEY is read from the environment; or pass api_key=...
embed_model = ForgeEmbedding(model="turbo")

vector = embed_model.get_text_embedding("the quick brown fox")
query = embed_model.get_query_embedding("fast animal")
batch = embed_model.get_text_embedding_batch(["doc one", "doc two"])

As the global embed model

from llama_index.core import Settings, VectorStoreIndex, Document
from llama_index.embeddings.forge import ForgeEmbedding

Settings.embed_model = ForgeEmbedding(model="pro")
index = VectorStoreIndex.from_documents([Document(text="hello world")])

Async

vec = await embed_model.aget_text_embedding("doc")
q = await embed_model.aget_query_embedding("a search query")

Matryoshka (shorter vectors)

embed_model = ForgeEmbedding(model="turbo", dimensions=256)  # re-normalized 256-d vectors

Configuration

Arg	Default	Notes
model	"turbo"	turbo | pro | ultra (stored as model_name)
api_key	FORGE_API_KEY env	get one at dash.voxell.ai
base_url	https://api.voxell.ai
dimensions	None	Matryoshka truncation, e.g. 256
timeout	30.0	seconds

License

MIT © Voxell, Inc.