Primitive of the day · 2026-10-01

ollama-bge-m3-dense-head-only

Ollama's bge-m3 GGUF serves ONLY the dense 1024-d head - sparse and ColBERT heads require FlagEmbedding's HF weights (a separate ~2.3 GB download), and FlagEmbedding cannot load the Ollama GGUF.

When this applies

Planning hybrid (dense+sparse+colbert) retrieval with bge-m3 while assuming the existing Ollama pull covers it.

Preconditions

• Hybrid retrieval requires the sparse and/or colbert heads
• bge-m3 currently available only via Ollama

Do this

Install FlagEmbedding and let it download the HF bge-m3 weights; use BGEM3FlagModel.encode() to get all three heads in one forward pass; keep Ollama's bge-m3 for dense-only paths.

What you should see

dense_vecs + lexical_weights + colbert_vecs from a single encode call.

How it fails if ignored

Expecting sparse/colbert from Ollama's /api/embed returns dense-only (verified: embeddings dim-1024, no sparse, no colbert); duplicate model downloads surprise disk budgets.

Do not use when

NOT relevant if only dense embeddings are needed (Ollama /api/embed suffices).

Kind: gotcha-fix. Part of the skill Bge-m3 triple vector qdrant. Free to reuse in your own agent skills.