Skip to main content

​
The Problem

RAG (Retrieval Augmented Generation) pipelines are the backbone of modern AI applications. From customer support chatbots to internal knowledge bases, every AI company is building some form of RAG. But here’s the reality: RAG pipelines require multiple components working in perfect harmony:
  • Data ingestion - Getting your documents into the system
  • Embedding generation - Converting text to vectors
  • Vector storage - Storing and indexing embeddings
  • Retrieval API - Querying relevant documents
  • LLM integration - Generating responses
When any component changes, the entire chain needs to adapt. Traditional approaches make this a nightmare.

​
The Traditional Approach

Timeline: 2-3 weeks (if nothing goes wrong)

​
Step 1: Research (2-3 days)

Choose your vector database. Pinecone? Weaviate? Qdrant? Milvus? Chroma? Each has different APIs, pricing models, and operational characteristics. Read the comparisons. Watch the YouTube videos. Ask on Discord.

​
Step 2: Deploy Infrastructure (2-3 days)

Set up hosting for your chosen database. Configure networking. Set up authentication. Handle TLS certificates. Debug why connections are timing out.

​
Step 3: Configure Embedding Pipeline (2-3 days)

Choose an embedding model (OpenAI? Cohere? Open source?). Set up inference infrastructure or API connections. Figure out batching and rate limits. Handle retries when the API is flaky.

​
Step 4: Wire Everything Together (2-3 days)

Connect your vector database to the embedding service. Write the glue code. Handle error cases. Figure out why documents are being embedded twice.

​
Step 5: Build the Retrieval API (2-3 days)

Create an API layer with authentication. Implement query preprocessing. Add relevance scoring. Handle the edge cases.

​
Step 6: Set Up Monitoring (1-2 days)

Configure alerting for when things break. Set up dashboards. Figure out what metrics actually matter.

​
Step 7: Maintain It Forever

Your data source schema changes? Manually update the ingestion pipeline. Embedding model gets deprecated? Rewrite the embedding service. Vector database needs a version upgrade? Schedule the migration. Each change ripples through every component. Total: 2-3 weeks of setup, then ongoing maintenance burden.

​
With pragma-os

Timeline: 15 minutes

​
Step 1: Browse the Store

Find the resources you need. Vector storage, embedding pipelines, and retrieval APIs are all available as declarative resources.

​
Step 2: Define Your Pipeline

# storage.yaml - Your document storage
provider: gcp
resource: storage
name: documents
config:
  location: EU
  storage_class: STANDARD

# embeddings.yaml - Your vector storage
provider: gcp
resource: bigquery-dataset
name: embeddings
depends_on:
  - gcp/storage/documents
config:
  location: EU

# retrieval.yaml - Your query API (coming soon)
provider: gcp
resource: cloud-run
name: retrieval-api
depends_on:
  - gcp/bigquery-dataset/embeddings
config:
  region: europe-west4

​
Step 3: Apply and Watch

pragma resources apply .
pragma-os provisions everything and establishes reactive dependencies between components.

​
Reactive Dependencies in Action

This is where pragma-os shines. Traditional pipelines break when things change. pragma-os pipelines adapt. When your data source schema changes:
  1. The document storage resource detects the change
  2. The embedding pipeline automatically adjusts to the new schema
  3. The vector index rebuilds with updated embeddings
  4. The retrieval API reflects the new structure
All automatically. No manual intervention. No 2am pages. Your RAG pipeline stays current with your data, not stuck in the state it was when you first deployed it.

​
What’s Coming

The examples above show the vision. Today, pragma-os provides the foundational GCP resources (Cloud Storage, BigQuery, Cloud Run) that serve as building blocks. Full RAG-specific resources are our next priority:
  • Vector database resources (Pinecone, Weaviate integrations)
  • Embedding pipeline resources with model selection
  • Pre-built retrieval patterns with best practices baked in
The goal: one-click RAG deployment where you specify your data source and pragma-os handles everything else.

​
Why This Matters

AI teams shouldn’t spend weeks on infrastructure plumbing. They should spend that time on what makes their product unique: the prompts, the user experience, the domain-specific logic. pragma-os handles the undifferentiated heavy lifting so you can focus on building something remarkable.

Get Started

Try pragma-os with your first resource in 5 minutes.