I've got a sort-of agent running in #ClaudeCode that I really love but I want to depend less on the #Anthropic ecosystem. I don't mind running foundation models, but I want to be able to change up what models I am using.

#OpenClaw is certainly an option. But a bit "Wild West" right now.

#HermesAgent has been interesting. I think my first stab at it could have gone better. But then I leaped in without really reading any of the docs. I'm thinking I might spin up another one just for the purpose of going through the docs and taking the time to learn it, concept by concept, before more thoughtfully designing the real agent I want.

The Claude Code agent that I'm using has a few things going for it that I'd like to port over (or improve upon):

• Molecular note keeping partner using #ObsidianMD. I've been at this awhile so my use is pretty sophisticated and I've had a hard time getting smaller local models to play along well.
• Agent memory based on #LadybugDB, tracking promises, semantic spacetime vertices & edges... This has been pretty cool and I think I want to recreate it (better) on #ArcadeDB.
• Also using #LightRAG as my hybrid RAG platform. I'd want to migrate this to #ArcadeDB, as well.
• I think I'm going to create more sophisticated MCP tools that are more thoughtfully designed for context management and containment of tool proliferation.
• I'd love to get to a place where I could use a Hermes Agent harness for most of my work, local models for a lot of smaller work (I've only got a 16GB VRAM GPU) and foundation models (or #OpenRouter) for the heavy lifting.

#AI #AIAgents

Một người dùng đang tối ưu hóa hệ thống RAG cục bộ bằng cách chia tải giữa GPU 5090 (LLM) và 5070Ti (embedding), chuyển sang kiến trúc song song với vLLM hoặc llama.cpp để tăng tốc độ xử lý. Họ cũng cân nhắc nâng cấp lên Qwen 30B và thảo luận về việc dùng Docker để quản lý mô hình. Câu hỏi trọng tâm: phân bổ GPU có hợp lý? vLLM hay llama.cpp tốt hơn? Nên dùng mô hình coder thay vì instruct? #AILLM #GPUTips #RAG #LightRAG #Optimization #CôngNghệAI #ThếHệGPU

https://www.reddit.com/r/LocalLLaMA/co

Использование графов знаний при разработке RAG-систем

Привет, Habr! На связи Александр Сулейкин, Роман Бабенко и Даниил Бутнев. Подготовили совместную статью по теме использования графов знаний при разработке RAG-систем. В рамках статьи рассказываем про основные проблемы традиционных RAG-систем, даем обзор основных открытых проектов графов знаний GraphRAG, показываем и даем краткое описание архитектуры таких систем, а также рассказываем про практическое использование графов знаний на примере трех областей - медицины, метрологии и стандартизации.

https://habr.com/ru/articles/908890/

#graphrag #LightRAG #RAG #ai #llm

#開源分享 中國北郵和港大開源了一個新的基於圖的RAG系統：LightRAG，進階版GraphRAG
解決了上下文感知能力不足的問題，在響應效率、成本和對新資訊的適應性上比GraphRAG有升

特點：
1、引入了圖結構到文本索引和檢索過程中
2、雙層檢索框架，低級和高級檢索，可適應不同類型查詢
3、將圖結構與向量表示結合，提高實體和關係的檢索效率
4、具備增量更新能力，可以及時整合新數據，確保系統在動態環境中保持有效性和響應性

專案地址： github.com/HKUDS/LightRAG

#RAG #GraphRAG #LightRAG