AI Glossary
Plain-language definitions of AI, LLM, RAG, agent, and deployment terms — each with osFoundry’s angle and links to the relevant feature.
Models & inference
- Base Model
A base model is a pre-trained LLM before any task-specific fine-tuning or RLHF. osFoundry treats base models as the starting point for fine-tuning (LoRA / QLoR… - Context Window
The context window is the maximum number of tokens an LLM can process in one request (input + output combined). Modern models range from 4 K to 2 M tokens. osF… - Hot-swap LoRA
Hot-swap LoRA is the ability to switch between LoRA adapters on a deployed base model at inference time without reloading the model. osFoundry hot-swaps up to… - Large Language Model
A large language model (LLM) is a neural network trained on vast text corpora to predict the next token, producing fluent natural-language output. osFoundry ro… - LoRA
LoRA (Low-Rank Adaptation) fine-tunes only a small number of "adapter" parameters on top of a frozen base model, drastically reducing training cost. osFoundry… - Mixture of Experts
Mixture of Experts (MoE) is an LLM architecture where only a subset of parameters is activated per token, giving the capacity of a large model at the inference… - Parameters
Parameters are the learned weights of a neural network. osFoundry’s catalog labels every model with its parameter count (e.g. 7B, 70B) so you can match model s… - QLoRA
QLoRA is LoRA fine-tuning where the base model is quantised to 4-bit during training, cutting VRAM requirements roughly in half. osFoundry supports QLoRA as an… - Quantization
Quantization reduces an LLM’s memory footprint by storing weights in lower precision (4-bit / 5-bit / 6-bit) instead of full 16-bit floats. osFoundry quantises… - Self-hosting
Self-hosting means running an LLM on infrastructure you control — your laptop, your data centre, or a dedicated GPU you provision. osFoundry’s self-host runtim… - Temperature
Temperature controls the randomness of an LLM’s output — 0 is deterministic, higher values introduce variability. osFoundry exposes temperature per chat reques… - Tokenization
Tokenization splits raw text into the sub-word units (tokens) that LLMs actually process. Different models use different tokenisers; the same text counts as a… - Top-p (Nucleus) Sampling
Top-p sampling restricts the model to the smallest set of tokens whose cumulative probability exceeds p. osFoundry’s pipeline configs let you tune top-p per ag…
RAG & retrieval
- Chunking
Chunking splits long documents into smaller pieces that fit comfortably in an LLM context window and are independently retrievable. osFoundry auto-chunks notes… - Embedding
An embedding is a high-dimensional vector representation of text that captures semantic meaning. Similar texts have similar vectors. osFoundry’s default embedd… - HyDE (Hypothetical Document Embeddings)
HyDE is a retrieval technique where the model first generates a hypothetical answer to the query, then embeds that answer for retrieval (instead of embedding t… - Knowledge Graph
A knowledge graph is a structured representation of entities and the relationships between them. osFoundry continuously extracts a workspace knowledge graph fr… - MMR (Maximal Marginal Relevance)
MMR is a re-ranking technique that balances relevance to the query against diversity among the selected results. osFoundry’s RAG pipelines include MMR as a con… - Reranking
Reranking is a second-pass step in RAG that re-orders the top-k retrieved chunks by relevance to the query, using a smaller dedicated model. osFoundry’s RAG pi… - Retrieval
Retrieval is the step in RAG where relevant documents or chunks are pulled from a knowledge store based on the user’s query. osFoundry combines vector embeddin… - Retrieval-Augmented Generation
Retrieval-Augmented Generation (RAG) is the technique of fetching relevant context from a knowledge store at query time and including it in the LLM prompt. osF… - Semantic Search
Semantic search retrieves content by meaning rather than exact keyword match, using vector embeddings. osFoundry’s workspace-wide search combines semantic with… - Vector Database
A vector database stores embeddings and supports fast nearest-neighbour search. osFoundry uses pgvector inside each per-app Postgres — no separate vector DB to…
Agents
- Agent Orchestration
Agent orchestration is the pattern of coordinating multiple agents (or multiple LLM calls) to accomplish a complex task. osFoundry’s Maestro is itself an agent… - AI Agent
An AI agent is an LLM with the ability to call tools, take multi-step actions, and pursue a goal autonomously. osFoundry hosts persistent agents with sessions,… - Function Calling
Function calling is the API mechanism that lets an LLM emit a structured request to call a developer-defined function. It’s how tool use is wired at the API le… - MCP (Model Context Protocol)
Model Context Protocol (MCP) is an open standard for connecting tools and data sources to AI applications. Originally introduced by Anthropic, it’s now support… - Memory Group
A memory group is a shared, versioned knowledge store that one or more agents read from and (optionally) write to. osFoundry’s agents attach memory groups to t… - Multi-Agent System
A multi-agent system uses multiple specialised agents that delegate sub-tasks to each other. osFoundry’s agents call other agents as typed tools, enabling plan… - Plan Mode
Plan mode is a pattern where an agent proposes a step-by-step plan before executing, allowing a human to review and approve. osFoundry supports plan mode at th… - Tool Use
Tool use is an LLM’s ability to call external functions (web search, code execution, APIs, file operations) and incorporate the results into its response. osFo…
Privacy & deployment
- Audit Log
An audit log records every user action, agent run, tool call, and admin change for security review and compliance. osFoundry’s audit log is SIEM-exportable wit… - End-to-End Encryption
End-to-end encryption ensures only the endpoints (sender, receiver) can read message content — intermediaries (including the platform vendor) cannot. osFoundry… - Local-first
Local-first is a software architecture where the source of truth is the user’s device, with optional sync to the cloud. osFoundry’s desktop app is local-first… - No-leak LLM Mode
No-leak LLM mode is a workspace setting that disables every cloud route — every model call must resolve to a local or self-hosted endpoint. osFoundry’s no-leak… - On-device Inference
On-device inference runs an LLM directly on the user’s hardware (laptop, phone) without any network call. osFoundry’s built-in inference server runs open-weigh… - RBAC (Role-Based Access Control)
Role-Based Access Control assigns permissions to roles (admin / member / guest) rather than individuals. osFoundry workspaces have RBAC with a separate `can_ma… - SSO (Single Sign-On)
Single Sign-On lets users authenticate once with their identity provider (IdP) and access multiple applications without re-entering credentials. osFoundry supp… - VPC (Virtual Private Cloud)
A Virtual Private Cloud is an isolated network within a public cloud provider, giving you full control over IP ranges, routing, and access. osFoundry’s BYO-clo…
Pricing & business
- BYOK Cost Model
BYOK (Bring Your Own Key) pricing means the platform doesn’t resell AI provider tokens — the customer pays the provider directly with zero markup. osFoundry us… - Per-seat Pricing
Per-seat pricing charges a fixed monthly fee per user, regardless of actual usage. osFoundry deliberately doesn’t use per-seat pricing — usage-based scales fai… - Usage-based Pricing
Usage-based pricing charges customers in proportion to what they actually consume — per query, per GB, per second of compute — instead of a flat per-seat fee.…