Chat node
The Chat node answers a question with grounded context it retrieves
itself: the flow’s embedded documents (vector search), a live web search, and
any text files wired into its Files handle. It is the “chat with your data”
node.
Handles
| Handle | Purpose |
|---|---|
Question | The incoming question (a Start input, Prompt or text payload) |
Files | Text files to ground the answer on (Upload File nodes) |
Answer | { answer, sources } — the answer and its grounding sources |
Configuration
- Answer model — the language model that writes the answer.
- Instructions — an optional system prompt.
- Search embedded data — retrieve matching chunks from the flow’s vector
namespace (needs the flow to have embedded documents, see Embed Data). Set the embedding model, chunk count
(
Top K) and an optional rerank model. - Search the web — add a live web search (provider + model + depth).
- Files — wire text files into the
Fileshandle; their content becomes part of the context.
How it answers
The node retrieves context (embedded chunks → web result → wired files), caps the assembled context, and hands it to the answer model with your question. Every retrieval failure degrades to “no context” — the answer just has less grounding, the node does not fail. Only the final answering call can fail.
Test chat
The inspector has a Test chat box: ask one question and see one grounded answer, using the node’s current configuration — no saved history.
Full chat page
Each Chat node has a shareable chat page — click Open chat on the node (or visit the URL). A flow can hold several Chat nodes, so the page addresses exactly one by its node id (the “chat id”):
/<shortId>/chat/<nodeId> On the page you can chat back and forth; the conversation is saved, the answer streams in as it is generated, and the view auto-scrolls to the latest message.
- Public flows — anyone can open the page, but sending a question requires signing in (each asker pays their own tokens).
- Private flows — only the owner and members can view or ask.
- Each signed-in user gets their own conversation per chat node.
Output
The node returns:
{
"answer": "…",
"sources": [
{ "kind": "vector", "label": "a:0", "text": "…" },
{ "kind": "web", "text": "…" },
{ "kind": "file", "label": "notes.md", "text": "…" }
]
} Notes
- Requires a saved flow for embedded-data retrieval (the flow id is the vector namespace).
- Retrieval and answer costs are billed through the gateway’s usage tracking.
Related
- Embed Data — making documents searchable.
- Search Data — semantic retrieval over the index.
- Web Search — the live-search backend.
- Upload File — the files you can ground on.