Getting started
Core concepts and glossary
The words SentientOne uses: agents, knowledge, chunks, sources, keys, conversations, credits, traces and request ids, in one place.
Concept2 min readUpdated 23 October 2026
These are the terms you'll meet in the app, the API and this help centre.
Building agents
| Term | Meaning |
|---|---|
| Agent | A configured assistant: a name, a system prompt, a provider and model, generation settings, its knowledge and its MCP tools. Each agent has a UUID, its agent id. |
| System prompt | The instructions the agent follows in every conversation. Up to 100,000 characters. |
| Provider key (LLM key) | Your own OpenAI, Anthropic, Google Gemini or Groq key. Every reply runs on it. Stored encrypted with AES-256-GCM. |
| Template | A starting point for a new agent. The 15 templates set the name, description and system prompt only. |
| Version | A snapshot written each time you save a change to the tracked settings or the prompt. You can apply an older version. |
| Playground | A chat in the dashboard for testing an agent, with latency, tokens, sources and tool calls shown beside each reply. |
Knowledge
| Term | Meaning |
|---|---|
| Source | One thing you added to an agent's knowledge: a document, a crawled website page, a snippet or a Q&A pair. |
| Chunk | A piece of a source, about 400 words with 40 words of overlap, stored with its embedding and full-text index. Snippets are kept whole; a Q&A pair is stored as two chunks. |
| Embedding | A 1536-number vector made with OpenAI text-embedding-3-small, used for search by meaning. |
| Hybrid search | A vector search and a Postgres full-text search run together, merged with reciprocal rank fusion (k = 60). |
| Rerank | Cohere rerank-v3.5 re-scores the fused candidates against the question; the top 5 go to the model. |
| Sources event | The part of the stream that lists the passages the model was given, with titles, URLs or file names and scores. |
Shipping and calling
| Term | Meaning |
|---|---|
Secret key (sk-so-) | Your account's API key, for servers. It can call any of your agents. See API keys. |
Publishable key (pk-so-) | A widget key for one agent, limited to the websites you allow. Safe to put in a web page. See Publishable keys. |
| Conversation | A thread of messages with one agent. The API returns a conversation_id; send it back to continue. |
| Server-sent events (SSE) | The streaming format replies use: data: lines, each holding a JSON object with a type. |
| MCP server | A Model Context Protocol server that exposes your systems as tools. Agents call its tools over HTTP. |
| Workflow | A graph of steps (blocks) on a canvas, run from a webhook or the API once published. |
Operating
| Term | Meaning |
|---|---|
| Credit | The unit you prepay. 1 credit per agent reply; 5 per document or website page; 1 per snippet or Q&A pair. Credits never expire. |
| Trace | The record of one request as OpenTelemetry spans: auth, retrieval, model rounds, tool calls and tokens. Kept for 90 days. Its id is the trace_id. |
| Request id | A correlation id carried in X-Request-Id. Send your own or the API makes one; it appears in error bodies and logs. See Request ids. |
| Rate limit | Requests per minute per agent, 60 by default for new agents, counted in a fixed one-minute window. |
| Answer cache | Exact repeats of the first question of a new conversation, for an agent without tools, answered from a 15-minute cache. Still 1 credit. |
Still stuck?
Email us with what you tried and any request id. A person replies.