SkillRAE (CUHK): Retrieval-Augmented Execution beats all agent skill retrieval baselines. Offline skill graph + dual retrieval + context compilation resolves LLM skill dependencies.
TGS-RAG claims 80% fewer tokens than LightRAG on HotpotQA at equal or better accuracy — a potential Pareto improvement for graph-augmented RAG systems.
SubQ's 12M-token context frontier model is a step-change above the 1–2M plateau and forces a fundamental rethink of RAG vs in-context architectures.
CUHK Hangzhou paper formalizes that skill MDs and RAG stores are lookup tables, not learning: retrieval needs combinatorial coverage; weight updates generalize from O(N) examples.
Memori hits 81.95% LoCoMo accuracy at just 1,294 tokens/query — 67% smaller prompts than Zep, 20x cheaper than full-context — with MCP server and multi-agent attribution model.
Skill-RAG predicts LLM failure via hidden-state probing, retrieves only when needed, and routes failure types to specialized skills — beating RAG benchmarks on efficiency and accuracy.
OpenAI is deprecating text-embedding-3-small, prompting calls to open-source the model so trillions of indexed tokens remain queryable after the closed-source model is retired.
An analysis of context engineering patterns emerging from 50 production AI deployments — covering RAG architectures, knowledge graph integration, multi-layer memory systems, and the shift from prompt engineering to structured context pipelines.
How leading organizations combine knowledge graphs with LLMs to build AI systems that reason over structured relationships — covering GraphRAG architectures, entity resolution, and the emerging graph-native context engineering paradigm.
Curated AI insights — sent when there's something worth your inbox.