Canonical Concepts

A sophisticated architecture for Retrieval-Augmented Generation that incorporates multi-stage processing like pre-retrieval query rewriting, hybrid search, second-stage re-ranking, and post-retrieval context compression.

The foundational developer stacks, containerized runtimes, and orchestration frameworks required to build, host, evaluate, and scale autonomous AI agents.

A technical mechanism for retaining, retrieving, and managing state across an AI agent's interactions over both short and long time horizons, simulating cognitive architectures.

An autonomous retrieval paradigm where an LLM agent uses tools to formulate search strategies, evaluate retrieved passages, and issue follow-up queries iteratively until sufficient context is gathered.

Okapi BM25 is a classic, probabilistic ranking function used in information retrieval to score document relevance based on exact keyword matching, term frequency, and inverse document frequency.

Techniques for dividing long documents into smaller, discrete text blocks to optimize embedding quality and retrieval precision within a vector database for RAG applications.

The dynamic allocation and management of fixed token limits across different prompt components, such as system instructions, retrieved RAG passages, conversation history, and tool outputs.

The practice of storing pre-computed Key-Value (KV) states of prompt tokens on model servers to eliminate redundant processing and drastically reduce latency and cost for repetitive context.

Techniques for condensing long text prompts, retrieved RAG passages, or conversation histories into fewer tokens while preserving the core semantic meaning and factual information.

The overarching discipline of designing, structuring, retrieving, budgeting, and dynamically injecting the optimal state and information into a Large Language Model's prompt window.

The process of dynamically populating prompt templates with fresh, relevant external data fetched from APIs, databases, or web scrapers at runtime.

The runtime administration of prompt windows, encompassing the dynamic control of token usage, truncation strategies, memory eviction, and state updates.

Algorithmic strategies for dropping older or less relevant tokens when a generated prompt exceeds the strict bounds of a models context window.

The maximum continuous sequence of tokens that a language model can accept, process, and attend to during a single inference pass.

A robust Retrieval-Augmented Generation strategy that dynamically evaluates the relevance of retrieved documents and triggers fallback actions, such as web searches or query rewrites, when confidence is low.

Neural network architectures that compute deep, bidirectional attention across query and passage tokens simultaneously to produce highly accurate relevance scores.

A semantic search paradigm utilizing dense neural embeddings, where all vector dimensions contain continuous real numbers representing the latent meaning of text.

The foundational pipeline for extracting, parsing, and converting complex unstructured document layouts into clean, structured context streams for AI ingestion.

The programmatic, real-time insertion of runtime variables, user metadata, retrieved passages, and tool outputs into system prompt templates.

The algorithmic process of identifying, disambiguating, and linking disparate records that refer to the same real-world entity across different datasets.

A temporal log of an agent's past experiences, interactions, observations, and execution steps recorded in strictly chronological order.

Retrieval-Augmented Generation that leverages knowledge graphs, structural hierarchies, and community summaries to answer complex, global, and multi-document queries.

A unified retrieval approach that combines sparse lexical matching and dense vector search, typically merged via scoring algorithms like Reciprocal Rank Fusion (RRF).

A structured, explicit network representation of real-world entities, concepts, and relationships stored as nodes and edges in a graph database.

Traditional keyword-based text search that matches exact words, substrings, or tokenized stems using inverted index structures.

Externalized, persistent storage that allows AI agents to recall facts, user preferences, and historical context across separate execution sessions and vast timeframes.

The automated background process of summarizing, pruning, and structurally transferring transient working memory into persistent long-term storage.

Algorithmic recency weighting and garbage collection strategies applied to historical memory records to prioritize current context and manage state size.

An open standard protocol (MCP) that unifies and securely standardizes how AI models and applications connect to external data repositories, databases, and tool servers.

Vector representations that map multiple data types—such as text, images, and audio—into a single, shared latent vector space for cross-modal similarity search.

The foundational Retrieval-Augmented Generation (RAG) workflow involving basic text chunking, embedding generation, vector similarity search, and direct prompt insertion.

Needle-in-a-Haystack (NIAH) is an evaluation methodology used to measure a Large Language Model's ability to retrieve a specific, targeted fact placed randomly within a massive context window.

A formal specification of conceptual classes, properties, constraints, and relationships within a specific domain, serving as a structural blueprint for knowledge representation.

A form of agentic memory that stores execution rules, tool usage routines, and multi-step workflow strategies, dictating how an AI agent performs tasks.

A second-stage retrieval process where a high-precision model (often a cross-encoder) re-evaluates and re-orders a preliminary list of candidate passages to optimize relevance.

Reciprocal Rank Fusion (RRF) is an algorithmic technique for combining multiple ranked search result lists (e.g., dense vector search and sparse keyword search) into a single, unified ranking.

An architectural pattern that dynamically retrieves relevant external information and injects it into an LLM's prompt window to ground text generation in factual, up-to-date knowledge.

A framework where large language models are fine-tuned to output special reflection tokens, allowing them to dynamically evaluate context relevance and self-correct generations.

A caching system that evaluates user queries based on vector similarity, returning previously generated responses or retrieval results for semantically identical questions.

An abstraction layer that maps complex, technical database schemas and raw data into consistent, business-friendly concepts accessible to both humans and AI agents.

Structured, generalized knowledge, facts, and concepts abstracted away from specific temporal interaction events, forming a durable knowledge base for AI agents.

Search techniques that interpret user intent and conceptual meaning rather than relying solely on literal string or keyword matching.

Retrieval techniques that rely on high-dimensional vectors where most values are zero, optimizing for keyword precision, rare term matching, and exact identifier lookup.

Protocols and message schemas for transferring task state, memory slices, and execution objectives between parent orchestrators and specialized child agents.

A hierarchical classification system of concepts organized into formal parent-child (is-a) relationships for structuring domain knowledge.

The capability of Large Language Models to generate structured JSON parameters to invoke external functions, APIs, or database queries.

The process of deconstructing raw documents like text, HTML, PDFs, and slide decks into machine-readable semantic elements.

Retrieval mechanism based on calculating mathematical distances between high-dimensional vector embeddings of queries and documents.

The automated extraction, cleaning, and formatting of web pages into structured markdown context streams tailored for Large Language Models.

The immediate, active context window space currently available for a Large Language Model's inference during a single interaction turn.