Glossary

Definitions of key terms used throughout the Orchid documentation.

A quick-reference dictionary of terms used across the Orchid documentation. Each entry links to the page where the concept is explained in depth.

Term	Definition
Agent	A named, YAML-configured unit of intelligence that handles a specific domain. Each agent has a description, a set of tools, and an optional custom class. See Agents.
Agentic loop	The inner reasoning cycle where an LLM issues `tool_calls`, the framework executes them, and the results are fed back until the LLM produces a final answer. See Agents.
AGENTS.md	A per-package Markdown file containing AI-specific context: structure, rules, patterns, and pitfalls.
Bearer token	An HTTP `Authorization: Bearer <token>` credential. Orchid resolves it to an `OrchidAuthContext` at the API entry point via `OrchidIdentityResolver`. See OAuth & Auth.
Boilerplate	The minimal code needed to wire a new consumer agent into Orchid — typically a YAML entry plus an optional Python subclass of `OrchidAgent`.
DCR (Dynamic Client Registration)	RFC 7591 — the protocol Orchid uses in `oauth` MCP auth mode to register a new OAuth client with a server's authorisation server at runtime, without pre-provisioned credentials. See OAuth & Auth.
Embedding	A fixed-length numeric vector that encodes the semantic meaning of a text chunk. Orchid uses embeddings to index and retrieve documents in the vector store (Qdrant or ChromaDB). See Embeddings.
Hierarchical RAG	Orchid's five-level scoping system (root → tenant → user → chat → agent) that controls which vector collections an agent can read from. See Hierarchical RAG.
Bloom	The execution layer of Pollen + Bloom. Triggers match incoming Signals and run them through the LangGraph supervisor as `JobRun`s. See Pollen + Bloom.
`LLMProvider`	The abstract interface (`orchid_ai.core.llm`) for simple text completions. Injected by the graph builder; consumer agents call `self.summarise()` rather than depending on a concrete provider.
Mini-agent	A parallel sub-agent forked by an opt-in parent during a single turn. The parent decomposes the request, N minis run focused agentic loops in parallel, and an aggregator synthesises the outcomes. See Mini-Agents.
`OrchidAgent`	The base ABC (`orchid_ai.core.agent`) that every agent class extends. Provides `run()`, `summarise()`, `extract_conversation_history()`, and RAG helpers. See Agents.
`OrchidAuthContext`	A dataclass (`orchid_ai.core.state`) carrying `tenant_key`, `user_id`, and `bearer_header`. Resolved once at the API entry point and threaded through the graph. See OAuth & Auth.
`OrchidChatStorage`	The ABC (`orchid_ai.persistence.base`) for chat session and message persistence. The library ships a PostgreSQL implementation; consumers can provide alternatives. See Persistence.
`OrchidMCPClient`	The abstract MCP client (`orchid_ai.core.mcp`) combining `OrchidMCPToolCaller` and `OrchidMCPDiscoverable`. Handles capability discovery, caching, and tool invocation. See MCP Integration.
`OrchidRAGScope`	A frozen dataclass (`orchid_ai.rag.scopes`) that constructs typed, hierarchical filter objects for the vector store backend. Always use this — never pass raw `tenant_id` strings. See Hierarchical RAG.
Parse-once	The pattern of calling `extract_text()` exactly once per document and passing the result to both the prompt builder and `ingest_document(pre_extracted_text=...)`. Avoids double-parsing cost. See Document Parsing.
Pollen	The event-ingestion substrate of Pollen + Bloom. Ingests, persists, and dispatches immutable `Signal`s from webhooks, schedulers, agents, or custom producers. See Pollen + Bloom.
Passthrough auth	MCP auth mode where the user's bearer token is forwarded unchanged to the MCP server. The server must trust the same identity provider as the Orchid API. See MCP Integration.
Scope (RAG)	A level in the five-level hierarchy (root / tenant / user / chat / agent) that determines which vector collections are visible to a retrieval query. See Hierarchical RAG.
Sliding-window summarization	An opt-in supervisor feature that compresses older conversation turns into a summary, keeping only the most recent exchanges verbatim. Reduces token usage for long sessions. See Supervisor.
Supervisor	The LangGraph node that routes user messages to the correct agent, manages conversation history, and assembles the final response. Configurable via the `supervisor:` YAML section. See Supervisor.
Tool strategy	The algorithm used by `GenericAgent` to invoke tools during skill execution: `all` (parallel), `sequential`, or `llm_decides`. See Tool Strategies.
Vector store	A database optimised for approximate nearest-neighbour search over embedding vectors. Orchid supports Qdrant (full-featured, containerised) and ChromaDB (embedded, on-disk, zero-infra CLI default). See Hierarchical RAG.
YAML-first	Orchid's design principle that new agents, tool strategies, and parser registrations are added via YAML configuration rather than code changes, following the Open/Closed principle. See Configuration.