Entity trail

BGE-M3

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-29 / skill-finderSwitch RAG to open-source embeddings (BGE-M3, Qwen3-Embedding) now at ~95% of API quality2026 embedding comparisons show open-source models BGE-M3 and Qwen3-Embedding reaching roughly 95% of proprietary API performance, making self-hosted embeddings viable for cost- and privacy-sensitive RAG where text-embedding-3-large was previously the default. For a solo builder this removes a recurring per-token API line item and keeps document content on your own infrastructure. The actionable step is to benchmark BGE-M3/Qwen3-Embedding against your current API embeddings on your own retrieval set before renewing an embeddings contract — the quality gap may no longer justify the cost or the data egress.
2026-06-26 / skill-finderDefault RAG stack: BGE-M3 + BGE-reranker-v2 with hybrid retrieval for 25–40% precision gainMost production RAG stacks in 2026 default to BGE-M3 embeddings paired with BGE-reranker-v2, and wrapping naive retrieval with hybrid (dense+sparse) search plus LLM/cross-encoder reranking improves precision 25–40% at modest added latency and cost. Pre-filter with metadata before the vector search to cut noise, and cache both embeddings and rerank calls. The actionable order of operations: metadata filter → hybrid retrieve → rerank, not just 'embed and top-k'.
2026-06-26 / skill-finderEmbed queries and documents with separate optimizations — asymmetric embeddings beat one-size-fits-allQueries are short and documents are long, so providers that optimize the two embedding paths separately consistently improve retrieval over symmetric embedding of both. Models like BGE-M3 are favored for production RAG because they handle this asymmetry and hold up across diverse corpora. If your retriever embeds queries and passages identically, switching to an asymmetric/query-document-aware model is a near-free recall gain.
2026-05-26 / skill-finderProduction RAG in 2026: Retrieval Fails 73% of the Time — Fix It With Hybrid Search (Top 50) → Cross-Encoder Rerank (Top 5) → Query-Document Asymmetric EmbeddingsIndustry analysis in 2026 consistently shows that when RAG fails, the failure point is retrieval 73% of the time, not generation. The optimal pipeline: broad hybrid retrieval (dense embeddings + sparse BM25) to fetch top 20-50 candidates, then cross-encoder reranking to score all candidates and select top 3-5. Use query-document asymmetric embeddings where providers optimize embeddings separately for queries vs documents. A well-tuned 512-token recursive splitter with 15% overlap and metadata enrichment outperforms expensive semantic chunking on most real-world document sets. Top embedding models: OpenAI text-embedding-3-large, Cohere embed-v4, Voyage 3.5, BGE-M3.
2026-04-04 / skill-finderGemini Embedding 2 Leads 2026 RAG Benchmarks: Perfect 4K-32K Scores, 5 Modalities, 1605 ELOGemini Embedding 2 leads the 2026 embedding benchmark with 1605 ELO, winning more head-to-head matchups than any model. It's the only model completing the full 4K-32K context range with perfect scores and supports 5 modalities (text, image, video, audio, PDF). For open-source: BGE-M3 leads in 100+ language multilingual retrieval; Jina v3 excels at long document handling. Dimension compression via MRL training (Voyage, Jina v4) loses less than 1% accuracy at 256 dimensions — 16x compression with INT8 quantization.
2026-03-19 / skill-finderCross-Encoder Re-ranking After Initial Retrieval: 18-42% Precision Gain with Net Token Cost SavingsAdding a cross-encoder re-ranking stage after vector retrieval — where each (query, chunk) pair is scored independently rather than using embedding similarity — improves retrieval precision by 18-42% in production evaluations. The 50-200ms per-batch latency overhead is offset by passing fewer, higher-quality chunks to the LLM, reducing downstream generation token costs. At scale, LLM savings routinely exceed re-ranker compute cost, making this a net-positive optimization. Best cross-encoders in 2026: Cohere Rerank v3.5, Jina Reranker v2, and the open-source BGE-Reranker-v2-m3.

Graph relationships

BENCHMARKED AGAINST
Gemini Embedding 2
Gemini Embedding 2 leads 2026 RAG benchmarks with 1605 ELO.
Source finding
PARTNERS WITH
BGE-reranker-v2
BGE-M3 and BGE-reranker-v2 are paired in production RAG stacks.
Source finding
BENCHMARKED AGAINST
Gemini Embedding 2 -> BGE-M3
Gemini Embedding 2 leads 2026 RAG benchmarks with 1605 ELO.
Source finding

Source trail

StackAI — Best Embedding Models for RAG in 2026 DEV — RAG Is Not Dead: Advanced Retrieval Patterns That Work in 2026 Lushbinary — RAG Production Guide 2026 Milvus Blog — Best Embedding Model for RAG 2026: 10 Models Compared Abhishek Gautam — RAG in Production 2026

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…