Architecture
Orchestration and agentic retrieval
A workflow graph that drives a knowledge layer: query planning, parallel subqueries over per-study sources, and synthesis with references.
What it shows
The orchestrator is a workflow graph with a fixed, checkpointed order. Callers and other agents reach it; agents come in through a governed MCP tool, under the same permissions as a person.
Retrieval is agentic: a model decomposes the question into subqueries, they run in parallel as keyword, vector or hybrid searches with semantic reranking, each over the index of one study, and a synthesis step merges the grounding with its citations.
Three shared services sit beside the graph. A determinism cache returns reviewed answers without a model call. A model router sends each step to the right model, with the version pinned. Audit receives every step, the retrieval activity log and the token counts.
This is a reference design for an agentic clinical-trials Q&A assistant. It is a design, not a description of a system built as shown.
Key decisions
- One index per study: isolation is structural, not a filter that can be forgotten.
- The effort of query planning is a setting per query class, because planning costs tokens and time.
- Models are pinned by version, and the version is part of the cache key and of the audit record.
- Other agents get one permissioned tool server, not a point-to-point integration each.