osFoundry / Platform
Pipeline stages
Define and version every stage of your AI pipeline — retrieve, rerank, generate, post-process.
What it is
Pipeline stages is the part of osFoundry Studio where you break an AI workflow into named, ordered steps instead of one opaque prompt. A typical retrieval-augmented pipeline runs retrieve → rerank → generate → post-process, and here each of those becomes a distinct stage you can configure, inspect, and change on its own. Retrieve pulls candidate context from your knowledge and data; rerank narrows it to what actually matters; generate calls a model; post-process cleans, validates, or reshapes the output before it leaves the pipeline. Because the steps are explicit, you can see exactly where an answer came from and which stage to fix when it is wrong.
Inside osFoundry, every stage is defined in Studio and runs through the same BYOK model layer as the rest of the workspace, so you choose the model per stage — a small local model for reranking, a cloud model for generation, whatever fits — and you pay your provider directly with no reseller markup. Stages are versioned: when you change a stage's prompt, model, or parameters, the previous definition is preserved, so you can compare versions, roll back a regression, and know which configuration produced a given result. Being a local-first desktop workspace, your pipeline definitions and the data they touch can stay on your device rather than passing through a hosted service.
This matters because most AI quality problems are stage problems. When retrieval brings back the wrong documents, no amount of prompt-tuning on the generate step fixes it. Separating and versioning the stages turns "the AI is wrong" into a specific, testable question about one step, and lets you improve a pipeline the way you improve any other engineered system: one change at a time, with a history you can trust.
Key capabilities
- Define an AI pipeline as ordered, named stages — retrieve, rerank, generate, post-process — instead of one monolithic prompt
- Version every stage independently, so you can compare configurations, roll back a regression, and see which definition produced a result
- Pick the model per stage with your own API keys — local, cloud, or self-hosted — and pay providers directly with no markup
- Isolate quality problems to a single step: know whether an issue is in retrieval, reranking, generation, or post-processing
- Keep pipeline definitions and their data local-first on your desktop rather than routing through a hosted service
Use cases
- RAG builder: separates retrieve and rerank from generate so a wrong answer can be traced to bad retrieval instead of blamed on the prompt
- ML engineer: versions each stage while tuning, comparing a new reranker or generation model against the previous definition before committing
- Privacy-conscious team: runs retrieval over local knowledge and keeps the whole pipeline definition on-device, calling cloud models only at the generate stage
Frequently asked questions
What are the stages in an osFoundry AI pipeline?
A common pattern is retrieve → rerank → generate → post-process: pull candidate context, narrow it to the most relevant, call a model, then validate or reshape the output. You define the stages your workflow needs and set the order; each one is configured on its own.
Can I use a different model for each stage?
Yes. osFoundry is bring-your-own-key, so each stage can call a different model — a local model for reranking, a cloud model for generation, a self-hosted one elsewhere — using your own API keys with no reseller markup.
What does versioning a stage actually give me?
When you change a stage's prompt, model, or parameters, the earlier definition is kept. You can compare versions, roll back if a change makes results worse, and tie a given output back to the exact stage configuration that produced it.