Foundry overview

Turn raw data into knowledge your agents can use. A visual pipeline for ingest, clean, chunk, enrich, and index.

Foundry is Nora’s data pipeline surface. Agents are only as good as the knowledge behind them — Foundry is where you build that knowledge.

Every Foundry pipeline follows the same six-stage shape:

Trigger → Source → Clean → Normalize → Enrich → Output

You compose each stage from blocks, wire them, and publish. The pipeline runs on a schedule, on demand, or in reaction to an upstream event.

When to reach for Foundry

• You have a folder of PDFs, spreadsheets, or web pages and you want your Agent to answer from them.
• You have a Notion, Google Drive, or Slack workspace and you want it kept in sync.
• You have raw database rows and you want them re-shaped, tagged, and indexed.
• You want to strip PII, normalize schemas, or add embeddings before storing.

The six stages

Stage	Purpose
Trigger	Kicks off a run. Manual, scheduled, watch-folder, or webhook.
Source	Pulls raw items from the outside world.
Clean	Removes noise, redacts PII, deduplicates.
Normalize	Chunks, parses, and reshapes into a consistent format.
Enrich	Adds embeddings, tags, entity extractions, language labels.
Output	Persists into a vector index, document store, or graph.

Skip any stage that doesn’t apply. The wiring is checked at publish so you can’t have Enrich without Normalize.

Pipelines vs. Flows

A Flow runs an Agent. A Foundry pipeline prepares data for Agents to use.

You don’t build agent behavior in Foundry. You don’t build data ingest in a Flow. The separation keeps each surface focused on one job.

Publishing a pipeline

Same model as Flows: Draft and Published. Publishing freezes a version, and scheduled runs use the frozen version. Rollback works the same way.

Where to go next

Triggers

How pipelines start — cron, watch, webhook, manual.

Connectors

Google Drive, Notion, Slack, S3, databases.

Chunk & parse

Break documents into retrievable pieces.

Run logs

Inspect a run — stage-by-stage.