Offline RAG starter: Ollama, FastAPI, and Docker without the cloud
VaultRAG is a local RAG boilerplate: PDFs, Ollama embeddings, a private index, and SSE answers — for clients who will not send documents to OpenAI.
Stuzhuk Lab — Chemistry of Code
Chemistry of Code
← All posts
Tag
rag
All blog posts with this tag.
Neural network quality metrics: how to know the model actually works
How to choose neural network quality metrics: classification, computer vision, OCR, LLM and RAG, calibration and production indicators — what to measure so the model truly works in the product.
ACL and multi-tenancy in RAG: filter permissions before retrieval, not after
Production RAG cannot leave authorization to the model. How to isolate tenants, apply access lists before retrieval, close hidden leak channels, and measure recall under tight filters.
RAG document ingestion: PDFs, tables, OCR, and versions without silent garbage
A production ingest pipeline for RAG: immutable originals, PDF and scan parsing, tables that keep their headers, parser quarantine, versioning, and idempotent replay.
How LLMs Find Knowledge: From Tokens and Transformers to RAG
How text becomes tokens and vectors, what a Transformer actually does, why training is not RAG, and how to grow a Markdown knowledge base into retrieval.
How LLMs Are Trained: Pretraining, Fine-Tuning, LoRA, RAG, and Agents
A practical map of LLM adaptation: what changes model weights, what plugs in knowledge and tools, and how to choose an approach without “fine-tuning on PDFs.”
Vector databases in production: the index and storage layer for semantic search
How to run a vector database as an index layer: ANN algorithms, filters, tenancy, operations, and when pgvector beats a dedicated store.
AI Evaluation Harness in 2026: Golden Sets, Regression Gates, and Release Criteria That Survive Production
Build an AI evaluation harness that connects golden datasets, retrieval and generation metrics, CI regression gates, shadow traffic, release controls, observability, and LLM gateway FinOps.
Production RAG Engineering in 2026: Chunking, Hybrid Search, Reranking, and Evals
A practical guide to production RAG architecture: ingestion, chunking, hybrid retrieval, reranking, evaluation, observability, security, and staged delivery for engineering teams.
Prompt Injection in Production 2026: Defending RAG, Tools, and Agents Against Untrusted Content
A production security model for prompt injection: treat retrieved documents, web pages, tool results, and MCP resources as hostile input; isolate context, constrain egress and tools, sandbox execution, and test layered defenses.
Chunking experiments for PDF, HTML, and knowledge bases
Run controlled chunking experiments while preserving document structure and citations.
How to build a golden dataset for RAG evaluation
Build a representative, versioned golden dataset for RAG retrieval and answer evaluation.
Hybrid search with BM25, vectors, and reciprocal-rank fusion
Combine BM25, vector retrieval, and reciprocal-rank fusion with measurable production guardrails.
Reranking in production RAG: quality, latency, and cost
Operate a RAG reranking stage with explicit quality, latency, and cost budgets.
Reference Architecture for an Enterprise AI Platform
A practical architecture that connects business problems, knowledge, RAG, agents, security, and observability into one enterprise AI platform.
Secure Enterprise AI Development: Releasing Change Without Losing Control
A practical lifecycle for changing enterprise AI safely: requirements, testing, security review, controlled release, rollback, and monitoring.
Enterprise AI Observability: Seeing the Whole System
How to operate enterprise AI with observability across infrastructure, RAG, answer quality, security, and business metrics.
Evaluating Enterprise AI: Metrics You Can Trust
A practical framework for measuring enterprise AI across retrieval, answers, security, and business outcomes.
Enterprise RAG Architecture: From Search to a Knowledge Platform
How to design scalable RAG for many data sources, millions of documents, and access-controlled enterprise knowledge.
Preparing Data for RAG: Turning Archives into a Knowledge Base
How to clean, describe, chunk, and safely index enterprise documents for reliable AI answers.
How to Select an Enterprise LLM: Criteria Over Rankings
How to select a language model for enterprise AI using scenarios, security, cost, and infrastructure requirements.
The First Enterprise AI Pilot: From Knowledge Map to Measurable Value
How to select a bounded, secure enterprise AI pilot, define success criteria, and prepare for responsible scaling.
The Enterprise Knowledge Map: From Information Chaos to AI Foundation
How to discover, classify, and safely connect enterprise knowledge for RAG and enterprise search.
ERP as a Knowledge Source, Not a Corporate AI Competitor
How to connect corporate AI to ERP while preserving data authority, access roles, and operational controls.
Why Enterprise AI Needs Agents
How agent systems connect language models to enterprise services and automate processes while keeping people in control.
When Fine-Tuning a Model Is Actually Needed
How to separate enterprise knowledge from model behavior and decide when fine-tuning is worth the operational cost.
Security as an Architectural Principle for Enterprise AI
Why data protection, access controls, answer traceability, and audit must shape enterprise AI architecture from day one.
Enterprise AI Infrastructure: Design for Workload, Not for GPUs
How to design reliable, scalable enterprise AI infrastructure from business scenarios, security constraints, and service-level requirements.
Vector Databases: The Foundation of Enterprise Semantic Search
How vector databases make enterprise knowledge searchable at scale while preserving source systems and access controls.
Why Embeddings Changed Enterprise Search
How vector representations retrieve enterprise knowledge by meaning while preserving exact search and security requirements.
Why Enterprises Need RAG Instead of LLM Memory
How RAG separates enterprise knowledge from the model, keeps documents current, and preserves verifiable, permission-aware access.