retrieval

"schema markup alone shifts precise information extraction from 16% to 54%. That's not a marginal gain — that's the difference between being cited and being invisible."

"RAG is an abbreviation for Retrieval-Augmented Generation, which in simple terms is a mechanism that searches external documents and then generates an answer."

"As the conversation grows longer and the history expands, the model's attention becomes scattered, leading to decreased compliance with constraints, increased costs, higher latency, and hitting context limits."

"I replaced LLM knowledge storage with a database and it works. Long live LLMs (without the Hallucinations)."

"For those working with Large Language Models (LLMs) in production, what kinds of questions should I expect? Specifically: System design: Do they ask you to design things like RAG pipelines or LLM-based applications?"

"LLMはRAGのように毎回ゼロから考えなくていい。人間の脳みたいに、一度理解したことを整理して蓄積すればいい"

"Pure semantic search didn't degrade because of scale per se; it started missing retrievals because the query and the target content used different vocabulary for the same concept. The fix was an index-first strategy — a lightweight topic-tagged index that narrows candidates before the NL query runs."

"The point of this paper is that text chunks are tagged with 'utility questions', connected as a graph, and then 'thematic memories' are extracted from the spectral decomposition of that graph."

"Now there is multi-agent orchestration, tool-calling, memory management, and autonomous task execution being shipped into production systems."

"3指標の一言まとめ Recall@K → 正解が「網に入ったか」（網羅性） MRR → 正解が「何位に来たか」（順位精度） キーワードヒット率 → 取得チャンクの「中身が揃っているか」（内容充実度）"

"I’ve wanted to like this model, so much. Thought it might replace Qwen 3.5 27b. Keep coming back to it and trying it every time there’s an update, hoping it will have improved."

"A harness is software that complements the input and output of the LLM, providing input in a format the LLM can understand, and appropriately processing the output from the LLM to return it to the user."

"Hybrid RAG uses both. The two search results are integrated using RRF (Reciprocal Rank Fusion). The greatest benefit is that documents that only hit in one of the searches can be rescued and surfaced."

"AI traffic is breaking traditional caching architectures, not just at the CDN layer, but all the way down to the database."

"By unifying PDFs, emails, and HTML into Markdown, it offers the advantages of making chunk splitting easier and stabilizing search accuracy."

"Token waste is not about reading too much. It is about reading the wrong things."

"Neural Networks As Hierarchical Associative Memory"

"This mechanism enables the LLM to determine which tools are needed, such as searching documents if document search is required, accessing the history database if correspondence history is needed, or registering tasks if task registration is necessary."

"Our experiments on a real industrial dataset demonstrate that the generated knowledge base provides an excellent starting point for accelerating RCA tasks and improving network resilience."

"By linking Gemini Gem and NotebookLM, we have entered an era where anyone can easily create a 'simple RAG' to minimize Hallucination and elicit accurate, evidence-based answers."

"RAG（Retrieval-Augmented Generation）の検索パイプラインは、ほとんどの場合こう組まれている: クエリ → ベクトル検索 → Top-K取得 → LLMに全部渡す この固定パイプラインこそが、RAGの精度を制限している元凶だった。"

"That moment is not a model failure. It is an architecture failure. And it is exactly the problem that Retrieval-Augmented Generation, or RAG, was designed to solve."

"Study finds that even the best AI with web search enabled exhibits fact-recognition 'Hallucination' in about 30% of cases."

"Context engineering involves deciding exactly what the model should see at any given moment. This goes beyond writing a good prompt; it includes managing system instructions, conversation history, retrieved documents, tool definitions, memory, intermediate steps, and execution traces."

"FlashRAG is a 'RAG all-in-one set' that combines 16 of the latest RAG algorithms and 38 datasets."

"AI Memory is designed to 'maintain the user's preferences, past dialogues, intentions, and the context of ongoing tasks, persistently (Persistent) across sessions and time.'"

"Agentic Search is, in a way, a more advanced version of RAG."

"They map text, images, audio, and video into the same embedding space, so a query in one modality can retrieve results from all the others."

"In your case, if you use a skill sheet as your knowledge base, you can answer questions like 'What is your experience with the OpenAI API?' while citing the relevant sections of the skill sheet."