Knowledge & output

An agent is only as good as what it knows and how cleanly it answers. Attach knowledge folders so the agent retrieves and cites your own content, give it memory so it carries context between sessions, and define a structured-output schema so its answers come back as predictable JSON.

Attaching knowledge folders

In step 4 of the builder you can attach one or more knowledge folders to an agent. Those folders become the agent's reference library: when it answers, it retrieves the most relevant passages and grounds its response in them, surfacing citations so you can trace every claim back to a source. This is retrieval-augmented generation (RAG).

Retrieval runs over the same pipeline as the rest of the platform. Documents you upload to a folder are extracted, chunked, and embedded into a pgvector index, and the agent retrieves by cosine similarity combined with keyword matching. Embedding uses a fixed priority of models — OpenAI text-embedding-3-small, then Google text-embedding-004, then Ollama / Ollama Cloud nomic-embed-text. See Knowledge Base for the full pipeline and supported file types.

• Multiple folders — attach several so a single agent can draw on, say, product docs and pricing at once.
• Per-folder AI-context toggle — a folder only feeds the agent when its AI-context toggle is on, giving you fine control over what the agent can see.
• Website context sources — folders can include crawled websites linked as AI sources, so an agent can answer from your public site or docs.

Memory

Where knowledge is your curated reference library, memory is the context an agent carries between sessions — the durable facts and preferences it should remember about a person or a task. It lets an agent stay consistent over time instead of starting cold every conversation.

Memory is governed by Memory Assignments, which control which context types the assistant's tools may access — things like a knowledge folder, a calendar event, a website, a workspace, or a user's own memory. You stay in control of exactly what context an agent can reach. See Settings for where memory and assignments are configured.

Structured-output schema

By default an agent answers in free-form text. Give it a structured-output schema — a set of typed fields — and its final response comes back as schema-conforming JSON instead. The agent fills in your fields rather than writing prose, which makes its output machine-readable and reliable.

This is what makes an agent safe to build on:

• In a workflow — downstream nodes can parse the agent's output by field name, because the shape is guaranteed. Drop an agent into a graph and wire {{agent.output.field}} into the next node.
• Over the API — calls return predictable fields you can map straight into your own system, no fragile text parsing.
• Consistency — the same request always returns the same shape, so integrations don't break when the wording changes.