Retrieval service (RAG) architecture
An answer endpoint over your own documents: vector store, retrieval step, generation step, cached responses.
Every resource, and what it costs.
Projections from August 2026 list prices for always-on resources. Connect an account and these become the figures your provider actually bills.
| Node | Type | What it is | Projected |
|---|---|---|---|
| Answer endpoint | Network | 10M requests | $10/mo |
| Vector store | Database | db.t4g.large + pgvector | $105/mo |
| Source documents | Storage | 1 TB standard | $12/mo |
| Answer cache | Database | cache.t4g.small | $25/mo |
| Question received | Trigger | HTTP request | — |
| Retrieve context | Agent step | Top-k over the vector store | billed per token |
| Answer with citations | Agent step | Grounded generation | billed per token |
| vector_search | Tool | Similarity search | — |
| Budget cap | Guardrail | Halts runs over $400/mo | enforced at runtime |
| ArcFoundry runtime | Runtime | — | — |
Agent steps are priced from provider-reported token usage once the agent runs, not estimated. Guardrails cost nothing and are the reason a runaway agent cannot.
Similar templates.
Support agent on live systems
A support agent that reads the orders database and writes back through an approval gate. Agent steps and the systems they touch on one graph.
Agent with enforced write control
The pattern for letting an agent touch production: every write passes an approval gate, every action lands in a hash-chained evidence log.
Document processing pipeline
Files land in storage, an agent extracts structured fields, a person approves anything low-confidence, and the result is written to the database.
Open Retrieval service (RAG) on the canvas.
It loads as an editable graph. Connect an account or instrument an agent and the projected figures above become measured ones.