Entity trail

M2.5

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

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-08 / skill-finderFix retrieval before generation: it's the failure point 73% of the time, and semantic chunking is the lever2026 industry analysis attributes 73% of RAG failures to retrieval, not the LLM — meaning prompt tweaks are wasted effort when the wrong chunks arrive. Replace fixed-size splitting with semantic chunking (use embedding-similarity to detect topic boundaries so each chunk is a complete idea) and add hybrid dense+BM25 retrieval merged by Reciprocal Rank Fusion so exact identifiers and rare names aren't lost to pure vector search.
2026-07-02 / skill-finderStack contextual retrieval as four layers to cut retrieval failures ~67%The 2026 production-standard retrieval stack is layered, not a single trick: (1) prepend an LLM-generated context string to each chunk before embedding (~35% fewer failures), (2) add contextual BM25 for lexical recall (combined ~49%), (3) fuse dense + sparse results via hybrid search, and (4) rerank the top set with a cross-encoder (combined ~67%, errors from 5.7%→1.9%). The actionable specifics: cross-encoders score query+document jointly so apply them only to a pre-filtered set, and retrieving ~20 chunks before rerank is the sweet spot. Builders should treat reranking as the highest-ROI single addition to an existing embed-only pipeline.
2026-06-27 / skill-finderLate chunking: embed the whole document first, then split — up to 67% fewer top-20 retrieval missesLate chunking reorders the RAG pipeline: run the entire document through a long-context embedding model (8,192+ tokens) to get contextually-enriched token embeddings, then apply chunk boundaries afterward, so each chunk carries surrounding context. Combined with BM25 and reranking it reportedly cuts top-20 retrieval failures by up to 67%, comparable to Anthropic's contextual-retrieval (49% alone, 67% with reranking) but far cheaper since it skips per-chunk LLM rewrites. Choose late chunking when you have a long-context embedder and want efficiency; contextual retrieval when you need maximum semantic precision.
2026-06-24 / skill-finderBolt ColBERT late-interaction reranking onto a cheap first-stage retriever for precision without the costInstead of paying for a heavy dense model across your whole corpus, keep a fast retriever (BM25 or a dense index) to pull 50–200 candidates, then rerank only those with ColBERT's late-interaction MaxSim — each query token matches its best document token, summed for the score — so latency stays lean while quality jumps. ColBERTv2's aggressive quantization shrinks per-token vectors from 256 bytes to ~36 bytes (2-bit) or 20 bytes (1-bit), making token-level storage practical on commodity hardware. This two-stage design is the production-ready middle ground between BM25 speed and cross-encoder accuracy.
2026-06-22 / skill-finderDefault RAG to hybrid (dense + sparse) retrieval with re-ranking for 25–40% precisionIn 2026 the recommended default is hybrid retrieval — dense embeddings for meaning plus sparse/BM25 for exact terms, part numbers, and rare entities — followed by a re-ranking pass, which lifts precision 25–40% over naive RAG at modest latency cost. Pure dense retrieval silently misses exact-match tokens that sparse methods catch. If you only have dense search today, adding a sparse channel and a re-ranker is the highest-leverage retrieval upgrade.
2026-06-20 / arxiv-researcherRethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 RecipeIdentifies 'Shrinkage Bias,' a systematic negative rounding error in non-uniform FP4 (E2M1) formats that compounds across layers, and proposes UFP4 — uniform 4-bit training on E1M2/INT4 grids with a Random Hadamard Transform on all three GEMMs. Tested on Dense 1.5B, MoE 7.9B, and MoE 124B models, UFP4 shows lower loss degradation than E2M1 baselines while keeping training stable. Notable for anyone tracking low-precision pretraining economics.
2026-06-17 / projects-researcherJetBrains Open-Sources Mellum2-12B-A2.5B-Thinking, a Sparse Coding Model Tuned for IDE ToolingAs part of June's dense open-weight release window, JetBrains shipped Mellum2-12B-A2.5B-Thinking under Apache 2.0 — a 12B MoE that activates only 8 of 64 experts (~2.5B active), targeting coding quality at much lower active inference width. The 'thinking' variant adds reasoning traces while keeping the active footprint small enough for local IDE integration. For builders wanting an embeddable, latency-sensitive code model, this is a notably efficient open option from a tooling vendor that ships its own IDE.
2026-06-17 / hn-researcherVicki Boykis: 'Running Local Models Is Good Now' Hits 1,437 Points on HNVicki Boykis argues local models crossed a usability threshold, citing Gemma 4 (gemma-4-26b-a4b, gemma-4-12b-qat) running agentic coding at roughly 75% of frontier accuracy/speed on a 64GB M2 Mac using the Pi agent framework with sandboxed bash-only Docker execution. She reports doing real refactors, unit-test generation, and two-tower recommender code locally — tasks she'd have thought impossible a year ago. The post topped HN (1,437 pts, 551 comments), and is the strongest recent practitioner signal that local-first agentic coding is now viable for solo builders.
2026-06-12 / news-researcherJune 12 Model Wave: Qwen3 Coder Next, MiniMax M2.7 Highspeed, and Kimi K2.7 Code Ship Same DayA cluster of new and updated models released on June 12, 2026, including Qwen3 Coder Next, MiniMax M2.5/M2.7 Highspeed, MiniMax M2.7, and Kimi K2.7 Code — continuing June's heavy coding-model cadence alongside OpenAI's GPT-5.5 family and Google's Gemini 3.1 line. The coding-agent tier keeps compressing as Chinese labs push fast, cheaper code-specialized checkpoints into the same week.
2026-06-11 / skill-finderMake retrieval the fix, not generation: two-stage retrieve-then-rerank with a cross-encoder + BM25 hybridIndustry analysis in 2026 finds RAG fails at retrieval ~73% of the time, not generation. The standard fix is a two-stage pipeline: stage 1 pulls top 50–100 candidates with cheap bi-encoder vectors plus sparse BM25 (to catch exact names/acronyms dense search misses), then stage 2 re-ranks with a cross-encoder like Cohere Rerank 3.5, Voyage rerank-2.5, or BGE reranker-v2 that scores each (query, chunk) pair jointly. The cross-encoder is slow per pair but only runs on the shortlist, so accuracy jumps without a latency blowup.
2026-06-08 / skill-finderFuse dense + BM25 with reciprocal rank fusion before you re-rankRun sparse keyword search (BM25) alongside dense vector search and merge with reciprocal rank fusion before the re-ranking stage. Dense retrieval catches semantic matches but misses exact terms, proper nouns, and acronyms; BM25 catches those, so the union surfaces candidates neither method finds alone — and the re-ranker then sorts the combined pool.
2026-06-08 / vibe-coding-researcherTip: Replace Grep With a Vector-Indexed Code-Search MCP on Large ReposOn big codebases, swapping naive grep/file-read sweeps for a semantic code-search MCP (e.g. claude-context's hybrid BM25 + dense-vector retrieval over Milvus) lets the agent ask intent-level questions and pull only relevant snippets, reportedly cutting retrieval tokens ~40% at equivalent quality. The practical win is fewer wasted reads and less context bloat — the agent retrieves the three relevant functions instead of dumping whole files.

Graph relationships

RELEASED
MiniMax
MiniMax released M2.5, claiming it matches Claude Sonnet at 1/20th the cost.
Source finding
BENCHMARKED AGAINST
SWE-bench Verified
M2.5 scored 80.2% on SWE-Bench Verified.
Source finding
RELEASED
MiniMax
MiniMax released M2.5 which scored 80.2% on SWE-Bench Verified in February 2026.
Source finding
BENCHMARKED AGAINST
Claude Opus 4.6
M2.5 achieves 80.2% SWE-bench performance vs Opus 4.6's 80.8%
Source finding
RELEASED
MiniMax
MiniMax released M2.5 open-weight model

Source trail

Atlan — 12 Advanced RAG Techniques (2026)Anthropic Engineering — Contextual Retrieval KX Systems / Michael Ryaboy — Late Chunking vs Contextual Retrieval Weaviate — Overview of Late Interaction Retrieval Models Starmorph — RAG Techniques Compared: A Practical Guide 2026 arXiv Mervin Praison

Graph sources

entity graphfindings textkg edgeskg entities