> ## Documentation Index > Fetch the complete documentation index at: https://docs.thinnest.ai/llms.txt > Use this file to discover all available pages before exploring further. # Knowledge Node > Look up a grounded answer from your agent's knowledge base at a specific point in the call, and branch on whether an answer was found. # Knowledge Node The **Knowledge** node looks up an answer from your agent's [knowledge base](/docs/knowledge) at the exact point you place it in the flow — then lets the call **branch on whether an answer was found**. It runs **instantly** (no caller turn), so the caller hears no dead air. **Why a node, when the agent already has a knowledge base?** Left to its own judgment, an agent *may* look something up — or it may answer from memory and get it wrong. The Knowledge node makes the lookup **guaranteed and on-script**: at this step, the answer always comes from your knowledge base, and you decide what happens when nothing is found. That determinism is the whole point of a workflow. ## When to use it "What are your hours / prices / return policy?" — answer from your own content instead of risking a made-up reply. Route one way when the knowledge base has an answer, and a graceful fallback (transfer, apologize, take a message) when it doesn't. ## Configure it Select the node on the canvas to open the **Inspector**. | Setting | What it controls | | ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Query** | What to search for. Defaults to the **caller's latest message** (`{{last_user_message}}`), so a Knowledge node placed right after a question "just works". You can also write your own query with `{{variable}}` interpolation. | | **Knowledge source** | Which source to search, picked from a dropdown. The default — **the agent's knowledge base (all sources)** — is the usual choice; pick a specific source only when you want to scope this node to one of them. | | **Store answer in variable** | The [variable](/docs/voice/workflows/transitions) that receives the retrieved answer text — e.g. `{{knowledge_result}}`. A later Conversation node reads it aloud. | | **Found flag → variable** | *(Optional)* A **boolean** variable set to `true`/`false` depending on whether anything was found. Use it to branch. | | **Max passages** | How many passages to pull from the knowledge base (default 5). | Declare the variables you reference (the answer variable and the found flag) in the **Variables** panel first, so the node has somewhere to store its results. ## Reading the answer back to the caller The Knowledge node doesn't speak — it **stores** the answer. The next **Conversation** node reads it by referencing the variable in its prompt, for example: > *"Answer the caller's question using this information: `{{knowledge_result}}`. > Keep it short and natural."* This route-then-speak pattern keeps the spoken reply grounded in your content while still sounding like the agent. ## Branch on whether an answer was found Give the node a **Found flag** variable (say `{{kb_found}}`), then draw two edges out of the Knowledge node: An [equation edge](/docs/voice/workflows/transitions) `{{kb_found}} == TRUE` → a Conversation node that reads `{{knowledge_result}}`. An **unconditional** edge → a graceful fallback (apologize and offer to take a message, or a [Transfer](/docs/voice/workflows/nodes#transfer) to a human). Equation edges are checked first, so the answer path wins when something was found and the call falls through to the fallback otherwise. ## Worked example — return policy On a Conversation node, the caller says *"What's your return policy?"* The flow routes to a **Knowledge** node. Its **Query** is left at the default (`{{last_user_message}}`), so it searches for exactly what the caller asked. The answer lands in `{{knowledge_result}}` and `{{kb_found}}` is set. Edge `{{kb_found}} == TRUE` → an **Answer** Conversation node whose prompt reads `{{knowledge_result}}` back to the caller. The unconditional edge → a **"let me take a message"** node when nothing was found. ## Pairs well with Global nodes Make the question-handling stretch a [Global node](/docs/voice/workflows/nodes#global-nodes) so the caller can ask a factual question **at any point** — the flow jumps to the Knowledge lookup, answers, then returns to whatever the caller was doing. ## Next steps Declare the answer and found-flag variables and write the equation edge that branches on them. See how Knowledge sits alongside Condition, API Request, and Tool. Run an integration action — SMS, calendar, CRM — at a specific step. Add the sources this node searches.