#RAG
297 related articles
RAG Meets Agent Memory: How Agentic GraphRAG Keeps Agents from Forgetting
Agentic GraphRAG unifies the RAG context layer and Agent memory layer into one graph, using knowledge, working memory, and decision memory layers to solve decision staleness after document changes. Agents can cite exact source versions, and affected decisions are automatically flagged when documents update.
Is Paid Independent RAG Evaluation Worth It? A Buyer's Real-World Experience Review
Starting from a Reddit question, this article breaks down the procurement motivations, price ranges, deliverables, channel paths, and decision impacts of paid independent RAG and chatbot evaluations, helping you judge whether this evaluation budget is worth spending.
The RAG Validation Trap: Why AI Hallucinations Persist in Production and a Four-Layer Architecture to Fix It
RAG is no anti-hallucination cure-all: a 'faithful summary' of a wrong retrieved document still passes evaluation. This article breaks down the RAG validation trap and presents a four-layer production architecture: tighter sampling, metadata filtering, deterministic validation, and human review.
RAG Enters the Production Deep End: Configuration Traps, Index Poisoning, and the Agent Memory Evaluation Crisis
This week's RAG research shifts to a production lens: an empirical study of 654 configuration failures, a topology poisoning attack on HNSW indexes, FuseRAG's 2.7x speedup, and the FEB-30 benchmark and SAFE-Mem governance scheme that expose agent memory evaluation blind spots.
EPUB Conversion Pitfalls: Chapter Order Can't Be Guessed—The Spine Is the Only Truth
Developer shares EPUB-to-Markdown pitfalls: chapter order must come from the spine, not alphabetical sorting or manifest; footnotes need inlining; EPUB2 and EPUB3 nav files differ and need a shim. Final output is structured Markdown for RAG pipelines.
Confidence Gating: When Should a Customer Service Agent Hand Off to a Human?
Customer service bots shouldn't optimize for tickets answered but for wrong answers sent. A breakdown of four confidence gates for RAG support agents: escalate when retrieval finds no evidence, mandate citations, verify citation authenticity, and let hard rules override thresholds—plus how to measure auto-send error rates and design golden test sets.
Building a Hindi Scanned-Document RAG Assistant: OCR and Hybrid Retrieval in Practice
Developer Abhishek shares lessons building a Hindi/English scanned-circular RAG assistant: OCR quality decides everything, hybrid retrieval and reranking boost results, and citations plus approval workflows make the product trustworthy.
RAG Retrieval Gone Wrong? The Real Problem Lurks in Chunk Boundaries
RAG returns irrelevant results? The problem is often chunk boundaries and metadata design, not embedding models. A practical look at Agent overuse, framework wars, RAG chunking, and observability.
RAG Hybrid Search: Why Vector Retrieval Misses TS-999 and a Simple Fix
Vector search excels at semantics but misses exact IDs like TS-999. This article explains RAG hybrid search: BM25 + vector retrieval, fused with RRF, plus a 20-line Python implementation, Anthropic's measured results, and a rollout roadmap.
Naive RAG Is Lying to You: 7 RAG Architectures That Survive Production
Why does Naive RAG fail in production? This article breaks down the five retrieval failure modes and systematically introduces seven RAG architectures that survive production: hybrid search and reranking, query transformation, contextual retrieval, RAPTOR, GraphRAG, Self-RAG/CRAG, and agentic/adaptive routing, with cost and selection decision rules.
RAG Hybrid Retrieval: Does Combining BM25 and Vector Search Improve Evidence Retrieval?
Comparing BM25, vector search, and RRF hybrid retrieval in RAG: their strengths and limits, how RRF can rank generic advice first, and metrics like Hit@5 and MRR@5.
Hand-Building an AI Customer Service Agent in Kotlin: Engineering Practices for RAG, MCP, and Multi-Agent Systems
An in-depth look at DSupport, a hand-built Kotlin Multiplatform AI customer service project: no Agent framework, using RAG relevance thresholds, MCP remote tools, multi-agent orchestration, and code-level guardrails to achieve an engineering architecture where models handle understanding and code handles guarantees.
Building a Company Brain in 3 Days: How Graph Memory Makes Every AI Answer Traceable
Developer Anshu built Recall in three days: a FalkorDB graph holds company causal knowledge so every AI answer is traceable, memory is per-person and versioned, and Agents wait for human approval. Graph retrieval beat flat RAG 24.667 to 19.251, with all missed points disclosed.
Treat the Context Window as a Budget: Token Allocation and Retrieval Strategies
The context window isn't an unlimited container—it's a budget to manage. Five budget lines, an 8,192-window arithmetic example, and bundle-based retrieval to avoid 'rules without exceptions' errors.
When Word Styles Can't Be Trusted: Practical Lessons from DOCX-to-Markdown Conversion
A developer shares lessons from DOCX-to-Markdown conversion: Word styles are often bypassed, requiring heuristic fallbacks to detect manually formatted headings—key for RAG and Agent parsing.
How to Choose a RAG Chunking Strategy: A Practical Guide from Fixed Splitting to Late Chunking
How to choose a RAG chunking strategy? This article analyzes the applicable scenarios and pitfalls of fixed splitting, recursive splitting, semantic chunking, structure-aware chunking, parent document retrieval, and late chunking, combining Chroma's research report and an ACL 2025 paper to help you avoid the common problem of RAG retrieving the answer but answering incompletely.
RAG Without Writing Code? An Offbeat Use of Gemini Workspace Integration
No vector database needed — Gemini's Google Workspace integration enables folder-scoped context retrieval and writing style imitation. This article breaks down this "user-level RAG" approach: its actual usage, token overhead mechanics, and key differences from standard RAG in retrieval principles.
RAG Systems: The LLM Isn't the Hard Part—Retrieval Quality Is
RAG failures often stem from retrieval quality, not model capability. This article analyzes retrieval relevance, chunking, hybrid search, multi-document issues, evaluation, and lifecycle management to reveal what's truly hard in RAG engineering.
Metrics First, Dashboard Second: A Guide to Choosing Metrics for LLM Evaluation
A polished LLM evaluation dashboard is useless if you pick the wrong metrics. This guide breaks down metric selection by feature type: F1 and AUC for classifiers, caution with perplexity for generators, faithfulness for RAG—with F1/AUC code examples on a labeled held-out set and a complete checklist.
From Substack Newsletters to a RAG Knowledge Base: A Clean Markdown Pipeline in Practice
A complete pipeline for building a RAG knowledge base from Substack newsletters: paginated archive API fetching, body scraping, targeted HTML-to-Markdown cleaning with turndown, heading-aware chunking, incremental updates, and respecting paywalls. Includes ready-made tools and code references.