Meet Abby
Abby is Parthenon's AI-powered outcomes research assistant — a context-aware, privacy-preserving assistant that lives inside every page of the platform and speaks the language of OHDSI.
Why Abby Exists
Outcomes research on the OMOP Common Data Model is powerful but complex. Researchers face a steep learning curve across vocabulary navigation, cohort expression logic, estimation methodology, and data quality interpretation. Traditional documentation helps, but it can't answer questions like:
- "What's the right approach for handling time-at-risk in a self-controlled case series?"
- "My propensity score distribution looks bimodal — what does that mean?"
- "How do I map this local drug code to a standard OMOP concept?"
These questions require contextual expertise — knowledge of both the methodology and the specific state of the user's work. Abby was built to provide that expertise in real time, grounded in peer-reviewed literature and official OHDSI documentation.
Design Philosophy
Abby was designed around four principles:
1. Privacy First
Abby runs entirely on-premises. The language model (MedGemma 1.5 via Ollama), the vector database (ChromaDB), and all embeddings are local services. No data ever leaves your infrastructure. This makes Abby suitable for air-gapped healthcare environments, HIPAA-regulated deployments, and organizations with strict data governance requirements.
2. Grounded in Evidence
Every response Abby generates is informed by a retrieval-augmented generation (RAG) pipeline that pulls relevant context from:
- 79,070 vectors of OHDSI research literature (2,258 peer-reviewed papers, the Book of OHDSI, HADES package documentation, and community forum Q&A)
- 46,271 vectors of Parthenon platform documentation
- Per-user conversation memory and shared FAQ
Abby doesn't hallucinate methodology — she cites it.
3. Page-Aware Context
Abby knows where you are in the platform. When you ask a question on the Cohort Builder page, she responds as a cohort expression expert. On the Population-Level Estimation page, she becomes a propensity score specialist. This page-aware persona system spans 22 distinct contexts across the entire platform.
4. Learns from Your Organization
Through conversation memory and FAQ auto-promotion, Abby builds institutional knowledge over time. Common questions about your specific CDM configuration, local data quirks, or organizational workflows get captured and reused — reducing repetitive support requests and onboarding friction.
The Name
Abby is named in the tradition of clinical research assistants — approachable, knowledgeable, always available. She's the colleague who's read every OHDSI paper and remembers every forum thread.
What Abby Can Do
| Capability | Description |
|---|---|
| Concept Search | Semantic vocabulary search using SapBERT clinical embeddings |
| Cohort Guidance | Help building inclusion/exclusion criteria and temporal logic |
| Method Selection | Compare estimation approaches, explain negative controls |
| Result Interpretation | Explain Achilles results, DQD findings, and analysis outputs |
| Variant Summarization | Interpret genomic variants with ClinVar context |
| Schema Mapping | Suggest OMOP concept mappings during data ingestion |
| Platform Navigation | Guide users through features and workflows |
| Literature Grounding | Answer methodology questions with references to OHDSI publications |
Quick Start
Abby is available on every page of Parthenon via the chat panel in the bottom-right corner. You can also find her in the Commons workspace as a dedicated channel (#ask-abby) or by mentioning @Abby in any conversation.
No configuration is required for basic use — Abby is active by default on all Parthenon deployments with the AI service running.