langchain-voxell 0.1.0

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

langchain-voxell

LangChain 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 langchain-voxell

Usage

from langchain_voxell import ForgeEmbeddings

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

doc_vectors = emb.embed_documents(["the quick brown fox", "lazy dog"])
query_vector = emb.embed_query("fast animal")

embed_query and embed_documents set the Forge input_type (query / document) for you.

Async

vectors = await emb.aembed_documents(["alpha", "beta"])
q = await emb.aembed_query("a search query")

Matryoshka (shorter vectors)

emb = ForgeEmbeddings(model="turbo", dimensions=256)  # re-normalized 256-d vectors

With a vector store

from langchain_community.vectorstores import FAISS
from langchain_voxell import ForgeEmbeddings

store = FAISS.from_texts(["doc one", "doc two"], ForgeEmbeddings(model="pro"))
hits = store.similarity_search("query", k=2)

Configuration

Arg	Default	Notes
model	"turbo"	turbo | pro | ultra
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.