Glossary Topics

Core Foundations

• Retrieval-Augmented Generation (RAG): Traditional RAG combines Large Language Models (LLMs) with external data retrieval systems to ground generated answers in company-approved content. This hybrid architecture supports highly accurate and context-rich responses rather than relying solely on pretrained model knowledge. Learn more
• Agentic RAG: An advanced evolution of RAG that integrates autonomous AI agents to reason, plan and break down complex queries in real time. The system dynamically adapts retrieval strategies and iteratively gathers context across multiple sources to deliver precise, trusted answers. Learn more
• Agentic Knowledge Layer: A centralized, shared infrastructure layer that handles ingestion, indexing, retrieval, evaluation and governance once for all enterprise AI applications. It eliminates duplicate pipelines and allows teams to swap underlying LLMs without needing to rebuild or reprocess any indexed data. Learn more
• RAG-as-a-Service: A scalable, API-first platform delivery model that allows developers and enterprises to easily integrate context-aware AI workflows via REST APIs and Software Development Kits (SDKs). It provides complete, modular and enterprise-ready RAG pipelines without requiring teams to build or manage complex underlying infrastructure. Learn more

Ingestion & Data Management

• Knowledge Box: The primary, governed container within the Agentic Knowledge Layer that logically isolates structured, semi-structured and unstructured resources. It combines raw files, metadata, concept relationship networks and role-based permissions into a single, highly queryable knowledge surface.
• Resource: Any individual piece of unstructured or semi-structured content—including PDFs, DOCX files, spreadsheets, videos, audio clips and images—uploaded to a Knowledge Box. Each resource is automatically normalized, enriched with metadata and annotated to make its underlying insights fully discoverable.
• Smart Ingestion: A continuous data-transformation pipeline that extracts text, parses complex layouts, segments content and tags access controls on raw enterprise assets. It converts messy, multi-format, unstructured content into highly retrievable, organized and clean structured knowledge.
• Sync Agents: Automated, continuous data connectors that monitor external cloud repositories like SharePoint, Google Drive and Dropbox to detect changes. They automatically synchronize new and updated content into the Knowledge Box without requiring manual ingestion scripts or scheduled IT jobs.
• Sync Service: A lightweight, cross-platform background utility running on Windows, macOS or Linux to monitor local directories. It securely pushes on-premises and local files into your designated Knowledge Box, keeping hybrid deployments continuously up to date.
• Semantic Chunking (Smart Segmentation): An intelligent text-splitting method that segments documents along logical, contextual boundaries like paragraphs, sentences or video timestamps. This keeps arguments contextually intact, dramatically increasing retrieval accuracy while preventing the broken payloads of fixed-character chunking.
• Metadata Enrichment & Entity Extraction: The automatic parsing of documents during ingestion to extract system/user tags and discover semantic relationships between named entities. This maps connections between people, contracts and concepts, allowing the system to easily handle complex, relationship-aware queries.

Hybrid Search & Orchestration

• Multilayer Indexing: A production-grade indexing strategy that simultaneously indexes and queries enterprise content across three distinct search dimensions. It combines semantic meaning, keyword precision and relational knowledge graphs to succeed where single-strategy search pipelines fail.
• Semantic Vector Search (Layer 1): A retrieval layer that maps document chunks as high-dimensional vectors to search for content based on underlying conceptual meaning rather than exact word matches. This excels at understanding user intent and context-dependent questions and handles multilingual queries across different vocabularies.
• Full-Text Search (Layer 2): A keyword-based, deterministic retrieval layer using BM25 ranking to score text blocks by exact term frequency and document-length normalization. It is essential for terminology-heavy lookups, allowing users to accurately retrieve precise terms like SKU numbers, regulatory codes and clauses.
• Knowledge-Graph Search (Layer 3): A highly specialized retrieval layer that traverses structured entity networks and semantic relationships discovered during data ingestion. This enables the RAG pipeline to answer complex questions by connecting disparate concepts and trace threads across multiple files.
• Reciprocal Rank Fusion (RRF): A mathematical fusion algorithm that normalizes and merges the incompatible relevance scores of vector, keyword and graph searches. It synthesizes these diverse scoring methods into a single, perfectly optimized and ranked list of results for the generator.
• Search Configuration: A stored, reusable definition of query parameters—including search weighting, metadata filters, rephrasing prompts and top-k limits. This allows multiple applications to apply completely different search logic on the same index without duplicating underlying data.
• Retrieval Agent: An autonomous orchestrator that plans and executes complex queries by analyzing questions and breaking them down into multiple sub-steps. It iteratively searches directories, gathers matching context and feeds a clean, unified payload to the LLM.
• Model Context Protocol (MCP): An open standard systematizing the secure, real-time connection between reasoning clients and external database retrieval layers. It allows assistants like Claude to securely query your Progress® Agentic RAG Knowledge Box mid-conversation without custom API pipelines. Learn more
• RAG Lab & Prompt Lab: Interactive, model-agnostic developer playgrounds built inside the platform to quickly test and refine RAG configurations. Teams can comparative-audit prompts, adjust retrieval weights and compare LLM performances before writing any custom code.

Evaluation & Metrics (REMi)

• REMi (RAG Evaluation Metrics Intelligence): A built-in, continuous evaluation framework that utilizes a dedicated, fine-tuned LLM to automatically score production-scale RAG outputs. It replaces slow manual spot checks by grading every single answer generated by live traffic to proactively catch retrieval failures. Learn more
• Answer Relevance: A core REMi metric that evaluates how directly, completely and effectively the generated AI response answers the user's specific query. Low scores in this metric flag that the model is generating evasive answers, highlighting a potential need for a different model. Learn more
• Context Relevance: A core REMi metric that measures whether the retrieved content blocks are genuinely relevant to the intent of the user's question. Low scores point directly to a retrieval configuration issue, prompting teams to adjust chunking, indexing or search thresholds. Learn more
• Groundedness: A critical REMi metric that measures the extent to which the generated AI answer is strictly supported by the retrieved document context. This acts as the primary defense against AI hallucinations, identifying when the LLM is introducing ungrounded outside claims. Learn more

Security & Economics

• Zero-Trust Data Access: A security framework that validates user roles and document permissions directly at the point of retrieval before sending data to an LLM. This facilitates corporate AI answers that strictly respect existing source-level permissions and reduce the risk of leaking sensitive files to unauthorized users.
• Tenant Isolation: A fundamental security boundary that logically isolates all accounts and distinct Knowledge Boxes within your organization. This helps keep proprietary business data private, strictly partitioned and compliant with SOC 2 Type II, ISO 27001 and GDPR.
• Nuclia Tokens: The unified utility billing metric that covers the entire managed RAG pipeline, including indexing, search, reranking and evaluation. Instead of billing for static work like uploading, it charges only at retrieval time, measuring when a query actually produces an answer.