Embed Data node
The Embed Data node embeds incoming text into vector form and upserts the vectors into the flow’s vector namespace (an Upstash Vector index named after the flow id). A Search Data node — or an Agent’s vector-search tool — can then retrieve them by semantic similarity.
Sources
- Input handle — the main payload flowing through the chain (wire it from the Start node).
- Documents handle — extra floating sources (e.g. an Upload or Prompt node).
Both resolve identically:
- A string passes verbatim.
- A
textupload contributes its content (txt / markdown / CSV). - Excel files (
.xlsx,.xls,.xlsm,.ods) are parsed into tab-separated text before chunking, so spreadsheet data is searchable too. - PDF files are extracted with
unpdf(a serverless-safe PDF.js wrapper). - JSON shapes are serialized; images, video and audio (no readable text) are skipped.
To embed a document, wire Start → Embed Data (its Input) so the node is
reachable — the engine walks from Start and never executes an unreachable node.
Configuration
- Model — the embedding model. Defaults to
openai/text-embedding-3-small. The Search Data node (and an Agent’s vector tool) must use the SAME model per flow, or queries return nothing. - Chunk size — target characters per chunk (default 8000, clamped 200–12000).
- Chunk overlap — characters shared between consecutive chunks (default 200), so context survives a split.
Chunking
chunkText splits recursively, preferring paragraph / line / sentence / word
seams over a hard mid-token cut. A single node stores at most 500 chunks; a
source that would exceed it fails with a clear message instead of spawning
unbounded (and billed) embeddings.
Output
The node returns { embedded } — the number of chunks now stored in the flow’s
namespace for this node. The card shows “N chunks embedded” after a run.
Notes
- Requires a saved flow (the flow id is the namespace).
- Requires
UPSTASH_VECTOR_REST_URLandUPSTASH_VECTOR_REST_TOKEN; otherwise the step fails with a clear message. - Chunk ids are deterministic (
<nodeId>:<index>), so a replayed step overwrites rather than duplicates.
Related
- Search Data — retrieving embedded documents.
- Agent — vector search as a tool.
- Upload File — the documents you embed.