Deployment terms

A library or toolkit for building AI agents — providing primitives for tool calling, planning, memory, and orchestration so you do not rebuild the agent loop from scratch.

The phenomenon where an AI system's behavior changes over time without explicit code changes — caused by model version updates, training data shifts, or vendor-side changes.

A time-boxed, scope-limited deployment of an AI agent against a real workflow to measure quality, cost, and adoption before broader rollout. The standard 2026 enterprise procurement pattern.

Running model inference asynchronously over a large batch of inputs, traded for latency. OpenAI/Anthropic batch APIs are typically 50% cheaper than sync calls.

A deployment pattern where you supply your own model API key to the agent — token costs are billed to you directly, the agent vendor charges only for the software.

The standard modern retrieval approach where queries and documents are encoded as dense embedding vectors and matched by similarity — distinct from sparse retrieval (BM25, keyword search).

AI that runs on the device where data is generated — phone, laptop, IoT, vehicle, factory floor — rather than in a remote data center. Trades model size for latency, privacy, and offline operation.

A retrieval technique that combines vector (semantic) search with keyword (lexical) search, fusing the scores to get higher precision than either alone. The 2026 production-grade default for RAG.

Spending more compute at inference (longer reasoning chains, multiple samples, search) to improve quality on hard problems — the architectural bet of 2025–2026 reasoning models.

A large language model running entirely on hardware you control — your laptop, your server, or your data center — with no calls to external APIs.

The infrastructure layer that hosts a model and exposes inference over HTTP — covering batching, scheduling, KV-cache management, and request routing.

AI tools that let non-engineers build agents, workflows, or applications via visual interfaces — drag-and-drop, prompts, or declarative configuration instead of writing code.

A deployment where the agent runs entirely on infrastructure the customer controls — no agent code or customer data leaves the customer's network.

An agent whose source code is publicly licensed (MIT, Apache, AGPL) — you can self-host, fork, and audit.

Running LLM inference inside your security perimeter (VPC, on-prem, confidential compute) so prompts and outputs never leave your control. Mandatory for regulated industries.

A document-splitting technique that uses embeddings to detect semantic boundaries — produces more coherent chunks for RAG than fixed-size chunking, improving retrieval quality.

A routing technique that uses embedding similarity to send each request to the right model, agent, or workflow — instead of brittle keyword rules or expensive LLM classifiers.

An LLM inference and programming framework optimized for structured generation, agent workloads, and complex prompting patterns — competitive with vLLM on throughput and faster on JSON/grammar-constrained output.

A capable LLM in the 1B–13B parameter range — trained to compete with frontier-quality on specific tasks while running on consumer hardware or at fraction-of-frontier cost.

Serving LLM outputs token-by-token as they're generated, typically over SSE or WebSocket — the default deployment pattern for any user-facing AI in 2026.

Total cost of ownership — the all-in cost of running an agent including subscription, token spend, ops time, and integration work.

A high-throughput open-source LLM inference engine — pioneered PagedAttention to manage KV cache like virtual memory, dramatically improving GPU utilization for serving open models.