Workflows, Deep Context & RAG Audit

LIVE

Workflow Modal

Multi-step variant analysis runner with an interactive chat panel. 50/50 split: steps on the left, Workflow Chat Panel on the right. ARIA narrates every step as it runs, so the reasoning is visible. Ask follow-up questions mid-run without losing context.

Deep Context Mode

Multi-shard knowledge retrieval and synthesis for questions that exceed a single-document context. Real-time shard status: pending · routing · budgeting · cost-gated · processing · synthesizing · complete · failed.

Explicit cost gate. Deep Context pauses before the expensive synthesis step and shows you the cost estimate. You approve or reject before heavy processing starts.

Medical Records mode

Extracts clinical terms from records with assertion, temporality, and negation markers. Reconciles Claude + NER + ontology database. Direct Export to Phenopacket from the findings view. Red-teamed with Gemini and GPT on negation and family-history handling.

RAG Audit Panel

Non-blocking inspector showing every retrieval: timestamp, algorithm (FTS5, LIKE, paper lookup), query, chunk count, similarity scores. Expand any entry for retrieved text, source badges (Curated / Trial / PMC), and match reasons. Auto-refresh about every 5 seconds when visible.

Why it exists

Trust in RAG-backed answers depends on auditing the retrieval. A visible, live inspector prevents "why did it say that?" from requiring a developer.