Node-based RAG Retrieval.
A Retrieval node you wire in like any other. It pulls relevant context from your own docs, codebase, or knowledge base straight into the chain at runtime — embed → search → top‑k → context.
Retrieval is just another node.
No separate RAG pipeline bolted onto your app. You drop a Retrieval node on the canvas and wire it between two others — the same way you place an Investigate or a Plan.
An upstream node hands it a question. It embeds the query, searches your indexed corpus, keeps the top‑k chunks, and writes them to shared Context — where Plan, Create, and Review read grounded source instead of guessing. It runs, reports state, and meters its own spend like every other node on the graph.
corpus → embed → search → top-k → write run.context
Run the Retrieval node.
One loop, every time it fires: embed → search → top‑k → context. Press Run and watch the corpus light up with matches, the top‑k rank by similarity, and the result land in shared Context.
query → vector
1536-dim
scan 24 chunks
cosine sim
keep k = 3
drop the rest
write
run.context
24 indexed chunks · idle
No data leaves your machine — embed, index and search all run locally.
Grounded chains, not guesses.
Retrieval as a node means every workflow can be grounded in your own truth — with the same wiring, state, and accounting as the rest of the canvas.
Wire it anywhere in the graph.
Put a Retrieval node before Plan to ground a strategy in your specs. Put one inside a Parallel branch so twenty agents each pull their own context. Feed its output through a Merge. It is a first-class node — branch it, gate it, retry it, cap its concurrency.
Your corpus never leaves.
Index and search run on your machine, against a self-hosted vector store. Pair it with local embedding models for an air-gap-friendly RAG that never ships a byte of your source to a third party.
Every chunk is traceable.
Retrieved context carries its source path and similarity score, so a Review node — or a human — can check exactly where an answer came from.
Counts its own spend.
Embedding calls are metered per node and roll up into the run total in Analytics — retrieval cost is never a black box.
k, model, threshold.
Set top-k, the embedding model, and a minimum similarity threshold per node — different parts of a chain can retrieve differently.
Wire it to the rest.
Skills-as-Nodes
Publish a skill; it becomes a node anyone can drop on the canvas.
openLocal Models
Ollama, llama.cpp, LM Studio — run your embedding model fully local.
openMCP + Integrations
Pull source material from Notion, Linear, and more straight into the index.
open