Glossary

Key terms used throughout the EpsimoAI platform.

Access Mode : Controls how an assistant can be accessed. Assistants can be private (project-scoped) or public (accessible by anyone with the assistant ID or via the public API).

Agent : A LangGraph-compiled graph that orchestrates LLM calls and tool usage. Agent types include chatbot (simple conversation), tools (function-calling agent), and retrieval (RAG-based agent).

Assistant : A saved configuration that defines an agent's behavior — which LLM to use, what tools are enabled, and the system prompt. Stored in the assistants table.

Checkpoint : A serialized snapshot of a LangGraph execution state. Stored in PostgreSQL, checkpoints allow conversations to be resumed and replayed. Each run creates one or more checkpoints.

Cognito Exchange : An authentication flow (RS256) designed for exchanging AWS Cognito tokens for EpsimoAI JWTs. Currently defined but not deployed in production.

Embedding : A dense vector representation of text, generated by an embedding model (e.g. OpenAI text-embedding-ada-002). Used for semantic similarity search in the vector store.

JWT (JSON Web Token) : The authentication token format used by EpsimoAI. Tokens are signed with HS256 using JWT_SECRET_KEY, valid for 30 days, and include the user ID, email, and project ID as claims.

LangGraph : A framework built on LangChain for creating stateful, multi-step AI agent workflows as directed graphs. EpsimoAI uses LangGraph to orchestrate tool-calling loops, streaming, and checkpointing.

MCP (Model Context Protocol) : An open protocol for connecting AI assistants to external tools and data sources. EpsimoAI supports MCP servers via both stdio and SSE transport, allowing assistants to use third-party tool servers.

Project : A workspace that groups assistants and threads under a single owner. Every user has a default main project created on first login. Projects enable logical separation of different use cases.

Public Token : A mechanism for accessing public assistants without user authentication. Public assistants can be invoked via the OpenAI-compatible API endpoint (/public/openai/v1/chat/completions) using the assistant ID.

Run : A single execution of an agent on a thread. A run takes input messages, invokes the LangGraph agent, streams the response, and saves the resulting checkpoint. Runs consume usage tokens from the user's thread_counter.

Thread : A conversation container that holds the message history between a user and an assistant. Threads are scoped to a project and optionally linked to a specific assistant.

Tool : A function that an agent can invoke during a run. Built-in tools include web search (DuckDuckGo, Tavily), retrieval (vector search), Wikipedia, ArXiv, PubMed, DALL-E, and MCP servers. Tools are configured per-assistant.

Vector Store : A PostgreSQL-based store (using the pgvector extension) that holds document embeddings for retrieval-augmented generation. Documents are chunked, embedded, and stored when files are ingested via the /ingest endpoint.