Skill

Bge-m3 triple vector qdrant

Use all three bge-m3 output heads (dense+sparse+colbert) in Qdrant named-vector collection with RRF fusion for maximum local hybrid recall

Primitives inside (7)

• bge-m3-similarity-floor-calibration calibration

  bge-m3 cosine similarities have a ~0.72 floor and short texts cluster - never use absolute similarity thresholds with it; rank-fuse (RRF) instead.

  When: Designing thresholds or ranking logic on bge-m3 embedding similarity scores.

  Register interest — 1 credit (≈ €0.10) — no payment

• ollama-bge-m3-dense-head-only gotcha-fix

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

  Register interest — 1 credit (≈ €0.10) — no payment

• qdrant-query-points-over-search gotcha-fix

  qdrant-client REMOVED client.search() in 1.18+ - write against query_points (available since ~1.10) so code survives client upgrades.

  When: Writing or maintaining Python Qdrant client code, or pinning qdrant-client versions.

  Register interest — 1 credit (≈ €0.10) — no payment

• qdrant-server-side-rrf-prefetch query-shape

  One-call server-side hybrid fusion in Qdrant: query_points with per-rail Prefetch entries + FusionQuery(Fusion.RRF) - keep client-side fusion when you need per-rail visibility or a custom rrf-k.

  When: Fusing multiple named-vector rails in Qdrant with minimal round trips.

  Register interest — 1 credit (≈ €0.10) — no payment

• qdrant-sparse-idf-modifier-skip gotcha-fix

  Leave Qdrant SparseVectorParams at defaults for model-learned sparse weights (bge-m3/SPLADE-style) - Modifier.IDF is for raw term counts and double-weights already-learned importance.

  When: Creating a Qdrant sparse-vector field for model-learned lexical weights.

  Register interest — 1 credit (≈ €0.10) — no payment

• qdrant-string-id-uuid5-mapping gotcha-fix

  Qdrant rejects string point ids - map arbitrary ids deterministically via uuid5(NAMESPACE_URL, id) so re-ingest upserts instead of duplicating, and keep the original id in payload.ext_id.

  When: Ingesting documents with human-readable string ids into a Qdrant collection.

  Register interest — 1 credit (≈ €0.10) — no payment

• rrf-rail-k-below-corpus-size calibration

  Set each retrieval rail's k well below corpus size - dense/colbert rails return candidates even at ~0 score, and rail-k >= corpus size makes RRF presence-counting swamp the rank signal.

  When: Tuning per-rail top-k for multi-rail retrieval fused with reciprocal rank fusion.

  Register interest — 1 credit (≈ €0.10) — no payment

Neighbour skills

Skills whose primitives are closest to this one (bge-m3 similarity):

• LightRAG local ollama graphrag
• Instructor pydantic extractor
• Build RAG reranker pipeline