Back to work

Context-aware AI Assistant

Connecting structured summaries and follow-up questions through shared, carefully scoped context.

Internship projectApplied AI integrationCoze Workflow / Agent / TypeScript

Overview

This application integration project connects history summaries and follow-up questions within DoctorAppDemo. A useful follow-up conversation needs access to the relevant information behind a summary. At the same time, that context must stay within the correct patient and consultation boundaries, and missing information must remain visible as missing.

The project uses Coze Workflow, an Agent, TypeScript, and a Gateway as parts of the application flow. Structured source information connects these pieces, giving the summary and the conversation a shared basis.

My contribution

My work covers integrating the history-summary and follow-up-question flows. The focus is on passing structured information, sharing the appropriate context between those flows, and keeping the patient and consultation scope explicit. Handling absent information belongs to that integration work, alongside getting the application components to communicate.

At a glance / Illustrative diagram
  1. Structured information
  2. Scoped shared context
  3. History summary
  4. Follow-up questions

System flow

The workflow starts with structured information, carries it into a scoped context, and connects a summary with subsequent questions. A follow-up question can then be considered alongside the information relevant to that consultation. The diagram shows these application relationships at a high level.

Engineering focus

Context sharing is useful only when its boundaries are clear. Important design concerns include keeping one consultation’s information separate from another and making gaps in the source material understandable. A response should preserve that uncertainty: a missing value is not a normal finding, and a concise summary should not conceal what the source leaves unknown.

Status & scope

This internship project focuses on application integration. Functional checks with limited source material do not establish clinical summary quality. It makes no claim of diagnostic ability or demonstrated clinical effectiveness.