Overview 🔍 link

BDIViz features an embedded LLM-powered assistant to help you validate, explain, and refine schema alignment candidates. The agent provides contextual insights, interactive feedback, and harmonization support throughout your data mapping workflow.

LLM Validation Panel 🤖 link

The LLM Validation Panel offers real-time, context-aware suggestions and explanations for each match candidate:

• Diagnosis Summary: Evaluates if a match is likely valid based on metadata, column names, and user history.
• Explanation Cards: Highlights key reasons for each suggestion (e.g., name similarity, value overlap).
  • Click to expand for more detail.
  • Provide feedback via 👍/👎 to refine model accuracy.
• Target Schema Metadata: Enriched descriptions pulled from sources such as GDC.

Harmonization Assistant 🤖 link

The Harmonization Assistant acts as your copilot for schema mapping and harmonization tasks, offering interactive, conversational support.

How to Access link

Example Q&A Interactions link

What is this candidate about? link

A: Ask the assistant for a summary or description of the selected candidate.
Prompt: What is this candidate about?

What are the potential value mappings between the selected source and target attribute? link

A: Request the assistant to show possible value mappings between the source and target attributes.
Prompt: What are the potential value mappings between the selected source and target attribute?

Does the candidate I selected make sense? Why? link

A: Ask the assistant to explain the reasoning behind a suggested match.
Prompt: Does the candidate I selected make sense? Why?

Can you suggest more candidates from the schema? link

A: Ask the assistant to search for additional candidates from the top recommendations or the entire target schema.
Prompt: Besides the current recommendations, show me more candidates related to tumor sites or organs.

Can you accept or reject candidates for me? link

A: Let the assistant accept or reject candidates based on chat history and its own suggestions.
Prompt: From your memory and our chat, can you accept the candidate that makes the most sense?

Can you remember custom notes or metadata for later? link

A: Add custom notes or metadata to the assistant’s memory for future reference.
Prompt: Remember this metadata for the source dataset: CCRCC is a common problem among...

Can you remind me what this task is about? link

A: Ask the assistant to recall the current task or context.
Prompt: Recall from your memory, what is the context of this task?

Can you filter or rematch nodes based on criteria? link

A: Start a new task or filter nodes based on clinical or semantic criteria.
Prompt: Rematch with nodes related to clinical criteria.

Can you reorder or rescore candidates? link

A: Request the assistant to reorder or rescore candidates.
Prompt: Reorder the candidates based on relevance to AJCC. Rescore and update them.

Can you help me create a custom matcher task? (unstable) link

A: Ask the assistant to help create new matcher tasks.
Prompt: Help me add a matcher using the rapidfuzz library.

Tip: The Harmonization Assistant learns from your feedback and memory, improving its recommendations over time. Use the thumbs up/down and memory features to tailor the assistant to your project needs.