BILLY DUVALLE

®2024

BILLY DUVALLE

®2024

OpenRAG

OpenRAG

OpenRAG is IBM's platform for building retrieval-augmented generation (RAG) apps: AI that answers questions from your own documents, websites, and data sources instead of guessing. It brings the full pipeline into one product, from ingesting and parsing knowledge to retrieving the right passages and generating answers that cite their sources. OpenRAG comes in two versions. The open-source edition lets developers self-host, bring their own models, and customize every step. The managed SaaS edition gives teams the same capabilities without running infrastructure. IBM acquired OpenRAG from DataStax, and I lead design across both versions, including its transition to SaaS.

Client

IBM

DELIVERABLES

OSS + SAAS SAAS TRANSITION ONBOARDING INGESTION PATTERNS CITATIONS EVALS

Year

2025–2026

Role

SENIOR PRODUCT DESIGNER

Design strategy

OpenRAG serves two markets that hire it for different jobs. Open-source developers self-host, bring their own models, and want to see and change every step of the pipeline. They judge a tool by how fast it runs locally and how little it hides. SaaS customers are teams that want grounded answers without running infrastructure. They judge a product by time to value, security, and whether they can trust what it returns. My strategy was one core experience with two entry points. Ingestion, retrieval, and citations work the same way in both versions. Developers get control surfaced where they need it. Managed teams get sensible defaults, guided setup, and governance. Because of that shared foundation, what the community builds carries over to the SaaS product, and moving from a local prototype to a hosted deployment doesn't mean relearning the tool.

Strategic outcomes

When IBM acquired OpenRAG from DataStax, it had to become two things at once: an open-source tool developers love and a SaaS product enterprises can buy. I design from inside the platform. I understand how documents move through Docling parsing, chunking, and embedding into OpenSearch, and how Langflow orchestrates retrieval and generation. That fluency let me design for what the system actually does, not what a mockup assumes, and let me spot where users lose trust: a silent ingestion failure, a chunk that splits a table, an answer with no traceable source. I work shoulder to shoulder with our engineers. We prototype against real pipelines, pressure-test designs against technical constraints early, and make trade-offs together instead of handing specs over a wall. That partnership is how a single experience for both open-source and SaaS became buildable rather than aspirational. The result is a platform with shared foundations: simplicity-first onboarding, one ingestion model for every knowledge source, and a trust layer of citations and evals that lets users prove an answer is right. OpenRAG now grows as one product serving two markets, designed and built by a team that speaks the same language.

Strategic outcomes

When IBM acquired OpenRAG from DataStax, it had to become two things at once: an open-source tool developers love and a SaaS product enterprises can buy. I design from inside the platform. I understand how documents move through Docling parsing, chunking, and embedding into OpenSearch, and how Langflow orchestrates retrieval and generation. That fluency let me design for what the system actually does, not what a mockup assumes, and let me spot where users lose trust: a silent ingestion failure, a chunk that splits a table, an answer with no traceable source. I work shoulder to shoulder with our engineers. We prototype against real pipelines, pressure-test designs against technical constraints early, and make trade-offs together instead of handing specs over a wall. That partnership is how a single experience for both open-source and SaaS became buildable rather than aspirational. The result is a platform with shared foundations: simplicity-first onboarding, one ingestion model for every knowledge source, and a trust layer of citations and evals that lets users prove an answer is right. OpenRAG now grows as one product serving two markets, designed and built by a team that speaks the same language.

When IBM acquired OpenRAG from DataStax, it had to become two things at once: an open-source tool developers love and a SaaS product enterprises can buy. I design from inside the platform. I understand how documents move through Docling parsing, chunking, and embedding into OpenSearch, and how Langflow orchestrates retrieval and generation. That fluency let me design for what the system actually does, not what a mockup assumes, and let me spot where users lose trust: a silent ingestion failure, a chunk that splits a table, an answer with no traceable source. I work shoulder to shoulder with our engineers. We prototype against real pipelines, pressure-test designs against technical constraints early, and make trade-offs together instead of handing specs over a wall. That partnership is how a single experience for both open-source and SaaS became buildable rather than aspirational. The result is a platform with shared foundations: simplicity-first onboarding, one ingestion model for every knowledge source, and a trust layer of citations and evals that lets users prove an answer is right. OpenRAG now grows as one product serving two markets, designed and built by a team that speaks the same language.

Design decisions

Onboarding to model providers. Selecting and configuring a provider took several rounds of prototyping. The deciding rule was simplicity over feature richness: developers reach a working setup first and configure later. Ingestion patterns and capabilities. Knowledge comes in as uploaded documents, crawled websites, and linked data sources, each with its own setup, progress, and failure modes. I designed one shared pattern for adding, tracking, and fixing anything that feeds the knowledge base, from bulk ingest that tracks each source to clear status and recovery when something doesn't load. Chunk-level citations. Answers link to the exact chunks they came from, so trust comes from being able to check. Evals v1. A way to tell whether a change made answers better, prototyped against the PRD with strict scope and validated with UX research.

‹ Context-aware components