AI agent
Software that reads a visitor's message, decides what to do, and acts (writes, calls a tool, escalates).
Different from a scripted chatbot in that it uses a language model for intent understanding and tool selection. See
what is an AI support agent for the deep dive.
BYO LLM
Bring Your Own Large Language Model. You provide the API key for the model; the platform never marks up tokens.
Architectural alternative to bundled AI where the platform resells tokens from a model it pays for. See
BYO LLM, explained.
Chunking
Cutting a document into smaller pieces (chunks) so retrieval can find precise sections without losing context.
Good chunking respects paragraph and heading boundaries and uses an overlap (usually 50-100 tokens) so important sentences aren't split mid-clause. ~800 tokens per chunk is a strong default for support content.
Citation
A reference in an AI's answer back to the source document the answer was drawn from.
The single highest-leverage thing for making AI support feel trustworthy. Visitors click the link once, see the source, and stop second-guessing.
Containment rate
Percentage of conversations the AI handled end-to-end without escalating to a human.
Common 60-75% on a well-grounded agent with clean knowledge. Higher isn't always better: aiming for 100% breaks the escalation seam and tanks CSAT.
CSAT
Customer Satisfaction Score. Typically a 1-5 or 1-10 visitor rating at the end of a conversation.
AI-augmented support tends to match or beat human-only CSAT on tier-1 questions and slightly trail on complex tickets. The escalation experience matters more than raw AI quality.
Embedding
A list of numbers (a vector) that captures the meaning of a piece of text.
Two embeddings close together in vector space mean the underlying texts are semantically similar. Used in RAG retrieval to find chunks relevant to a query.
Escalation policy
Rules describing who gets notified, in what order, and within what SLA, when the AI escalates a conversation to a human.
Most platforms support level-based policies (try level 1 first, page level 2 if no response in 30 seconds, etc.). OpenAgent ships with Slack and Google Chat webhook channels for the notification side.
Fine-tuning
Updating a model's weights with additional training data.
Different from RAG. Fine-tuning is expensive, slow to update, and ties you to a specific model. Most teams don't need it; RAG covers 90% of what people call “train the AI on my data.”
Hallucination
When an AI confidently states something that isn't true.
Reduced (not eliminated) by grounding the model in retrieved context, prompting it to refuse when unsure, and citing sources. Eliminated entirely only by humans verifying every answer, which defeats the point.
HITL (Human-in-the-loop)
An architecture where AI handles part of the workflow and a human is present for review, escalation, or final action.
The dominant pattern for production AI support: AI on the boring head of the queue, humans on the high-stakes tail. See
AI vs human ROI.
LLM
Large Language Model. The underlying transformer-based model (GPT, Gemini, Claude, etc.) that powers modern AI agents.
Generates text by predicting the next token given a context. Customer support uses LLMs in tandem with RAG for grounded answers.
OpenAI-compatible endpoint
An HTTP API that accepts requests in OpenAI's chat-completions format.
The de-facto standard. Self-hosted runtimes (vLLM, TGI, Ollama, LM Studio, Together, Groq) all speak it. Means any platform supporting BYO LLM via an OpenAI-compatible URL can talk to any of them.
pgvector
A PostgreSQL extension that adds a vector data type and similarity search indexes.
Lets RAG retrieval be a boring database query rather than a separate vector database with its own credentials and outage page. OpenAgent uses pgvector by default.
Prompt injection
An attack where a visitor types instructions that hijack the AI's behaviour.
Defended against with prompt isolation (user input never lands verbatim in the system prompt), input filtering, and tool permission boundaries. Not a solved problem; layered defenses matter.
RAG
Retrieval-Augmented Generation. Look up your data at request time and paste it into the model's prompt.
The standard pattern for grounding an LLM in your private data. See
what is RAG.
Resolution
A conversation marked as fully handled by the AI without human help.
The billing unit for hosted AI products like Intercom Fin (~$0.99 per resolution). BYO LLM platforms don't bill per resolution because token cost is already the unit economics.
Retrieval
The first half of RAG: finding the chunks of your knowledge base most likely to answer the visitor's question.
Typically done via embedding similarity in a vector database. Quality measured by metrics like Recall@k.
Shadow DOM
A browser feature that isolates a piece of UI from the host page's CSS and JavaScript.
OpenAgent's embeddable widget renders in a Shadow DOM so it never collides with your site's styling or breaks your page's layout. One script tag, no surprises.
System prompt
The instruction prepended to every AI conversation that sets the agent's persona, rules, and constraints.
Where you tell the model to answer only from context, refuse when unsure, cite sources, and match your tone of voice. Underrated leverage point; most teams under-invest here.
Tool call
A function the AI agent can invoke during a conversation: lookup_order, escalate_to_human, search_kb, etc.
Each tool has a name, a description, and a schema for its inputs. The model picks tools based on the conversation context. Tools are how an agent “does things” instead of just chatting.