Volium/snowflake-arctic-embed-m-long-q8_0.GGUF

license: mit tags:

• —gguf
• —feature-extraction
• —embeddings
• —sentence-similarity ---

snowflake-arctic-embed-m-long-q8_0.GGUF

GGUF build of Snowflake's Arctic Embed M Long embedding model, quantized to Q8_0 for use with llama.cpp / llama-cpp-python. nomic-bert architecture, ~0.1B params, 146 MB.

This is the embedding model behind two retrieval-only LightRAG harnesses — prebuilt corpora over official engine documentation (FAISS index + knowledge graph each) that you connect to whatever model or agent you already use:

• —[Unreal Engine 5 LightRAG harness](#) — retrieval over dev.epicgames.com
• —[Blender LightRAG harness](#) — retrieval over docs.blender.org

If you're using either of those repos, this is the exact model each one's FAISS index was built with, and you need this exact model (not a substitute) to embed queries against them correctly.

Files in this repo

Both .rar files are flat — the images sit at the archive's root, with no folder already inside either one. Each harness's own README has the exact extraction command and destination folder; using "Extract Here" instead of following those instructions will put the images one level shallower than the chunk data expects, and every image path will fail to resolve.

Quick use (Python)

from llama_cpp import Llama

model = Llama(model_path="snowflake-arctic-embed-m-long-q8_0.gguf", embedding=True)
vec = model.embed("How do I add a cube using Python?")

One thing worth knowing if you're embedding a query to search either FAISS index above (as opposed to indexing new documents): both harnesses were built following Snowflake Arctic Embed's documented asymmetric-retrieval convention — queries need this exact prefix prepended, indexed document text doesn't:

QUERY_PREFIX = "Represent this sentence for searching relevant passages: "
vec = model.embed(QUERY_PREFIX + "How do I add a cube using Python?")

Skipping this doesn't raise an error — it just quietly returns a vector in a slightly different space than the one the index was built in, so results are worse without being obviously wrong. Each harness's own retrieval code already does this correctly; it's only worth knowing if you're writing your own embedding calls against the raw .gguf file directly.

For CLI/server use (llama.cpp, Ollama, LM Studio, etc.), see the "Use this model" section above on this page — those work out of the box against this repo.

License

MIT (this model). The two harnesses above carry their own source-documentation licenses (Epic Games' documentation terms; Blender's CC-BY-SA) — check those separately before redistributing either corpus further.