AISummary Project

Physician-Focused AI Summaries

AISummary Copilot

Delivers concise, citation-backed patient insights in seconds so physicians can spend more time caring and less time scrolling charts.

AISummary is implemented as the Next.js route /patientPromptEnum/page.tsx. It renders physician-facing patient summaries by: (1) resolving URL params (siteId, patientId, userId, promptEnum, chatUri, etc.), (2) fetching Prompt and Category metadata via Refine’s DataProvider (which calls Azure Functions → Cosmos DB), (3) rendering one or more PatientPrompt widgets, and (4) posting the physician’s typed question or selected prompt(s) to the Python Flask AI service at chatUri (/v1/chat_patient). The page accumulates and scores ChatAnswer parts (via chatAudits) and displays content with categories and citations.

Technologies Used

Specifications

string | undefined

string | undefined (parsed to number in page)

string | undefined

string | undefined (disabled when interactive=false)

string | undefined — absolute URL for Flask /v1/chat_patient

boolean — from ?aiPatientIsPrimed=true

boolean — auto-run on first render

string — default 'summarizeComboPrompts'

string | undefined — name for parent postMessage loading events

string — seed error banner

boolean — default true; set false with ?interactive=false

useAuth() reads a bearer token from parent window via postMessage and injects it on all Refine/Azure calls (axios interceptor) and forwards it to chatUri when present.

Base for Azure Functions (ui/aicopilot/src/providers/data-provider).

Optional prefix (joined to API_URL).

If no auth and origin is not localhost, the page returns notFound().

v1/clients (resource: 'clients') — used to derive clientType for category filtering.
v1/categories (resource: 'categories') — select options filtered by clientType.
v1/prompts (resource: 'prompts') — fetches Prompt records by site/clientType/promptEnum/text.
v1/chatAudits (resource: 'chatAudits') — polled to retrieve score/validity for an answerId.

Simple eq filters mapped to query string (e.g., { field: 'siteId', operator: 'eq', value: siteId }).

/patientPromptEnum/page.tsx (client component)

PatientPrompt (no SSR, skeleton during load)
AISummaryModal (no SSR)
PromptRenderer (no SSR)
ModalTriggerTextField (no SSR)
PromptChips, BackToTopFab (no SSR)

useAISummaryParams — parse URL params.
useAISummaryInit — bootstrap: fetch combo prompts, honor fetchAnswerOnLoad, hydrate localStorage (model, guardrails, recent questions, initial prompt).
usePromptResources — get Prompt[] by enum/text via DataProvider.
useSubPrompts — expand a combo prompt's fullText by '<new_question>' into multiple questions.
useAnswer — orchestrate selection (enum vs free-text), post parent startLoading, clear state, then set prompts list.
usePromptVisibility — chip visibility rules based on content/citations/completion.
usePromptTitle — derive section title from first answer's markdown header if present.
useModalFocus — bottom sheet open/close and focus management.
useAuth, useOrigin — bearer and origin awareness.

POST {chatUri} (Flask /v1/chat_patient); also supports OPTIONS CORS preflight.

application/json

Bearer {auth} (if available)

string | null — forwarded bearer for downstream services (optional).

string

number | number[] — single or multi-patient; special case: -1 triggers generic chat flow.

number

string — the physician's typed question or one sub-prompt from a combo.

string — e.g., 'gemini2' (default from localStorage).

object — key/value placeholder inputs captured by PatientPrompt.

string[] | string — selected categories (from Settings menu or prompt defaults).

PatientRecord | PatientRecord[] | null — EHR payload if provided via postMessage.

optional 'text' | 'audio' — if 'audio', returns audio/mpeg.

number

string — possibly a fallback single blob if parts are empty or for errors.

string[] — each entry is usually JSON (stringified); UI attempts JSON.parse; fallback to plain text.

object — may include merged placeholder values the UI stores back into state.

Each part, if parsable JSON, conforms to UI ChatAnswer { content, categories: string[], citationsMap: Record<string,(string|number)[]>, score?, scoreReasoning? }.

SettingsMenu toggles 'isUsingGuardrails'.

For each returned answer part that has answerId, when scoreAnswer=true, the UI polls GET v1/chatAudits?answerId=… until a record exists; then copies score, scoreReasoning, valid, invalidReason.

If !valid OR score < 7 (MINIMUM_SCORE_THRESHOLD) and retryCount < 3 (MAX_RETRY_ON_LOW_SCORE), PatientPrompt re-calls chatUri for that prompt; otherwise continues. Defaults when scoring disabled: score=10 and valid=true.

When multiple prompts, PromptChips render; visibility changes once content+citations exist and prompt is completed.

AISummaryModal provides free-text ask, recent questions (localStorage), suggested combo prompts, model picker, categories multiselect, and guardrails toggle.

Parent ↔ IFrame integration: startLoading/stopLoading with { module }, fetchAnswer events carrying { promptEnum | question, data, model, patientId }.

Back-to-top FAB via useScrollTrigger on the prompts container.

HTTP 401/403/404/500 specific messages; 413/context-length guidance; generic catch-all lists troubleshooting steps.

page.patientPromptEnum.initialPromptEnum
page.patientPromptEnum.userFreeTextQuestions
page.patientPromptEnum.model
page.patientPromptEnum.isUsingGuardrails
page.patientPromptEnum.categories

Dynamic import of heavy client widgets; SSR disabled.
Single DataProvider instance; axios interceptor injects bearer.
Minimal prompt lookups via usePromptResources and useSubPrompts.
Polling chatAudits with 5s delay until record is present.
Merges response.context to persist dynamic placeholder values across retries.

{ answerId?, prompt?, content?, categories?, citationsMap?, score?, scoreReasoning? }

Minimal EHR snapshot interface used by UI.

{ id, promptEnum, text, fullText, title?, categories? }

{ queryText, patientID, siteID, userID, model?, category?, categoryAttributes?, referenceID?, auth?, subcategories?, currentDateTime?, questions?, filters?, context?, data?, format?, noCache?, nudges? }

{ answerId, content, parts: string[], context?: object }

GET v1/clients?siteId=…
GET v1/categories?clientType=…
GET v1/prompts?siteId=…&promptEnum=… OR text=…
POST v1/chat_patient (chatUri from URL param)
GET v1/chatAudits?answerId=…

Block Diagram

graph TD
  subgraph Browser[Physician Browser]
    U[/URL with siteId, patientId, userId, chatUri/]
    Page[Next.js /patientPromptEnum]
  end

  U --> Page
  subgraph UI[Next.js App aicopilot]
    Page --> Params[useAISummaryParams]
    Page --> Init[useAISummaryInit]
    Page --> PR[usePromptResources]
    Page --> Sub[useSubPrompts]
    Page --> Ans[useAnswer]
    Page --> Chips[PromptChips]
    Page --> Modal[AISummaryModal]
    Page --> Renderer[PromptRenderer]
    Renderer --> PP[PatientPrompt]
  end

  subgraph AF[Azure Functions TypeScript]
    Clients[/GET v1/clients/]:::fn
    Cats[/GET v1/categories/]:::fn
    Prompts[/GET v1/prompts/]:::fn
    Audits[/GET v1/chatAudits/]:::fn
    DB[(Cosmos DB)]:::db
    Clients --> DB
    Cats --> DB
    Prompts --> DB
    Audits --> DB
  end

  subgraph PY[AI Chat Service Flask]
    Chat[/POST v1/chat_patient/]:::fn
    LC[LangChain + LLM]
    Chat --> LC
  end

  PR -- load prompts --> Prompts
  Page -- clientType lookup --> Clients
  Page -- category options --> Cats
  PP -- POST question/context/data --> Chat
  Chat -- ChatAnswer(parts,context) --> PP
  PP -- poll score --> Audits

  classDef fn fill:#eef,stroke:#335,stroke-width:1px;
  classDef db fill:#efe,stroke:#393,stroke-width:1px;

graph TD subgraph Browser[Physician Browser] U[/URL with siteId, patientId, userId, chatUri/] Page[Next.js /patientPromptEnum] end U --> Page subgraph UI[Next.js App aicopilot] Page --> Params[useAISummaryParams] Page --> Init[useAISummaryInit] Page --> PR[usePromptResources] Page --> Sub[useSubPrompts] Page --> Ans[useAnswer] Page --> Chips[PromptChips] Page --> Modal[AISummaryModal] Page --> Renderer[PromptRenderer] Renderer --> PP[PatientPrompt] end subgraph AF[Azure Functions TypeScript] Clients[/GET v1/clients/]:::fn Cats[/GET v1/categories/]:::fn Prompts[/GET v1/prompts/]:::fn Audits[/GET v1/chatAudits/]:::fn DB[(Cosmos DB)]:::db Clients --> DB Cats --> DB Prompts --> DB Audits --> DB end subgraph PY[AI Chat Service Flask] Chat[/POST v1/chat_patient/]:::fn LC[LangChain + LLM] Chat --> LC end PR -- load prompts --> Prompts Page -- clientType lookup --> Clients Page -- category options --> Cats PP -- POST question/context/data --> Chat Chat -- ChatAnswer(parts,context) --> PP PP -- poll score --> Audits classDef fn fill:#eef,stroke:#335,stroke-width:1px; classDef db fill:#efe,stroke:#393,stroke-width:1px;

Component Diagram

graph TB
  A[/patientPromptEnum/page.tsx/]
  subgraph Hooks
    P1[useAISummaryParams]
    P2[useAISummaryInit]
    P3[usePromptResources]
    P4[useSubPrompts]
    P5[useAnswer]
    P6[usePromptVisibility]
    P7[usePromptTitle]
    P8[useModalFocus]
    P9[useAuth]
    P10[useOrigin]
  end

  subgraph UI Components
    C1[PromptRenderer]
    C2[PatientPrompt]
    C3[AISummaryModal]
    C4[ModalTriggerTextField]
    C5[PromptChips]
    C6[BackToTopFab]
    C7[Loading]
    C8[SettingsMenu]
  end

  A --> P1 & P2 & P3 & P4 & P5 & P6 & P7 & P8 & P9 & P10
  A --> C1 & C3 & C4 & C5 & C6 & C7 & C8
  C1 --> C2
  C2 -.->|onPromptFetching/onPromptFetched| A
  A -.->|postMessage start/stopLoading| ParentApp

graph TB A[/patientPromptEnum/page.tsx/] subgraph Hooks P1[useAISummaryParams] P2[useAISummaryInit] P3[usePromptResources] P4[useSubPrompts] P5[useAnswer] P6[usePromptVisibility] P7[usePromptTitle] P8[useModalFocus] P9[useAuth] P10[useOrigin] end subgraph UI Components C1[PromptRenderer] C2[PatientPrompt] C3[AISummaryModal] C4[ModalTriggerTextField] C5[PromptChips] C6[BackToTopFab] C7[Loading] C8[SettingsMenu] end A --> P1 & P2 & P3 & P4 & P5 & P6 & P7 & P8 & P9 & P10 A --> C1 & C3 & C4 & C5 & C6 & C7 & C8 C1 --> C2 C2 -.->|onPromptFetching/onPromptFetched| A A -.->|postMessage start/stopLoading| ParentApp

Sequence Diagram

sequenceDiagram
  autonumber
  actor Physician
  participant Parent as Parent App (optional)
  participant UI as Next.js /patientPromptEnum
  participant Hooks as use* hooks
  participant DP as Refine DataProvider
  participant AF as Azure Functions (v1/clients, v1/categories, v1/prompts, v1/chatAudits)
  participant DB as Cosmos DB
  participant PP as PatientPrompt
  participant PY as Flask aichat (/v1/chat_patient)
  participant LLM as LangChain + LLM

  Physician->>UI: Open with ?siteId&patientId&userId&chatUri[=&promptEnum|&interactive]
  UI->>Hooks: useAISummaryParams() parse flags
  Note over UI: If no auth and not localhost → notFound()
  UI->>DP: list('clients', filter: siteId)
  DP->>AF: GET /v1/clients?siteId=...
  AF->>DB: query Clients
  DB-->>AF: Client[] (clientType)
  AF-->>DP: Client[]
  UI->>DP: list('categories', filter: clientType)
  DP->>AF: GET /v1/categories?clientType=...
  AF->>DB: query Categories
  DB-->>AF: Category[]
  AF-->>DP: Category[]
  UI->>DP: list('prompts', filters: siteId + promptEnum or text)
  DP->>AF: GET /v1/prompts?siteId=...&promptEnum=...|text=...
  AF->>DB: query Prompts
  DB-->>AF: Prompt[]
  AF-->>DP: Prompt[]
  DP-->>UI: Prompt[]
  UI->>UI: useSubPrompts split '<new_question>' → prompts[]
  UI->>PP: Render for each prompt (or free-text)

  loop For each question
    PP->>Parent: postMessage startLoading{module}
    PP->>PY: POST { chatUri=/v1/chat_patient }\n{ siteID, patientID, userID, queryText, model, category[], context{}, data }
    PY->>LLM: Build chain and fetch
    LLM-->>PY: ChatAnswer(parts[], context)
    PY-->>PP: 200 { answerId, content, parts[], context }
    PP->>AF: Poll GET /v1/chatAudits?answerId=… (if guardrails on)
    AF->>DB: Find ChatAudit
    DB-->>AF: score, valid, reason
    AF-->>PP: ChatAudit
    alt score < 7 or !valid and retries < 3
      PP->>PY: retry POST /v1/chat_patient
    else success
      PP->>UI: onPromptFetched(prompt, parts)
      PP->>Parent: postMessage stopLoading{module}
    end
  end

  Physician->>UI: Adjust model/categories/guardrails, then ask again
  UI->>PP: Re-run with updated settings

sequenceDiagram autonumber actor Physician participant Parent as Parent App (optional) participant UI as Next.js /patientPromptEnum participant Hooks as use* hooks participant DP as Refine DataProvider participant AF as Azure Functions (v1/clients, v1/categories, v1/prompts, v1/chatAudits) participant DB as Cosmos DB participant PP as PatientPrompt participant PY as Flask aichat (/v1/chat_patient) participant LLM as LangChain + LLM Physician->>UI: Open with ?siteId&patientId&userId&chatUri[=&promptEnum|&interactive] UI->>Hooks: useAISummaryParams() parse flags Note over UI: If no auth and not localhost → notFound() UI->>DP: list('clients', filter: siteId) DP->>AF: GET /v1/clients?siteId=... AF->>DB: query Clients DB-->>AF: Client[] (clientType) AF-->>DP: Client[] UI->>DP: list('categories', filter: clientType) DP->>AF: GET /v1/categories?clientType=... AF->>DB: query Categories DB-->>AF: Category[] AF-->>DP: Category[] UI->>DP: list('prompts', filters: siteId + promptEnum or text) DP->>AF: GET /v1/prompts?siteId=...&promptEnum=...|text=... AF->>DB: query Prompts DB-->>AF: Prompt[] AF-->>DP: Prompt[] DP-->>UI: Prompt[] UI->>UI: useSubPrompts split '<new_question>' → prompts[] UI->>PP: Render for each prompt (or free-text) loop For each question PP->>Parent: postMessage startLoading{module} PP->>PY: POST { chatUri=/v1/chat_patient }\n{ siteID, patientID, userID, queryText, model, category[], context{}, data } PY->>LLM: Build chain and fetch LLM-->>PY: ChatAnswer(parts[], context) PY-->>PP: 200 { answerId, content, parts[], context } PP->>AF: Poll GET /v1/chatAudits?answerId=… (if guardrails on) AF->>DB: Find ChatAudit DB-->>AF: score, valid, reason AF-->>PP: ChatAudit alt score < 7 or !valid and retries < 3 PP->>PY: retry POST /v1/chat_patient else success PP->>UI: onPromptFetched(prompt, parts) PP->>Parent: postMessage stopLoading{module} end end Physician->>UI: Adjust model/categories/guardrails, then ask again UI->>PP: Re-run with updated settings

Physician-Focused AI Summaries

AISummary Copilot

Delivers concise, citation-backed patient insights in seconds so physicians can spend more time caring and less time scrolling charts.

Technologies Used

Specifications

Routing & URL Params (useAISummaryParams.ts)

params

params

siteId

siteId

patientId

patientId

userId

userId

promptEnumParam

promptEnumParam

chatUri

chatUri

aiPatientIsPrimed

aiPatientIsPrimed

fetchAnswerOnLoad

fetchAnswerOnLoad

comboPromptEnum

comboPromptEnum

moduleToStopLoadingForParam

moduleToStopLoadingForParam

initialErrorText

initialErrorText

interactiveMode

interactiveMode

Auth & Environment

Auth & Environment

auth

auth

env

env

API_URL

API_URL

API_ROOT

API_ROOT

originGate

originGate

Data Access (Refine DataProvider)

Data Access (Refine DataProvider)

resources

resources

filtering

filtering

Core Components & Hooks

Core Components & Hooks

page

page

dynamicImports

dynamicImports

hooks

hooks

AI Interface (PatientPrompt → chatUri)

AI Interface (PatientPrompt → chatUri)

http

http

headers

headers

Content-Type

Content-Type

Authorization

Authorization

requestBodyShape

requestBodyShape

auth

auth

siteID

siteID

patientID

patientID

userID

userID

queryText

queryText

model

model