diff --git a/LLMonFHIR/FHIR Display/FHIRMultipleResourceInterpreter.swift b/LLMonFHIR/FHIR Display/FHIRMultipleResourceInterpreter.swift index 79a4cf0..7b83a07 100644 --- a/LLMonFHIR/FHIR Display/FHIRMultipleResourceInterpreter.swift +++ b/LLMonFHIR/FHIR Display/FHIRMultipleResourceInterpreter.swift @@ -77,11 +77,15 @@ class FHIRMultipleResourceInterpreter: DefaultInitializable, Component, Observab private func systemPrompt(forResources resources: [FHIRResource]) -> Chat { - let allJSONResources = resources.map(\.compactJSONDescription).joined(separator: "\n") + var resourceCategories = String() + + for resource in resources { + resourceCategories += (resource.functionCallIdentifier + "\n") + } return Chat( role: .system, - content: Prompt.interpretMultipleResources.prompt.replacingOccurrences(of: Prompt.promptPlaceholder, with: allJSONResources) + content: Prompt.interpretMultipleResources.prompt.replacingOccurrences(of: Prompt.promptPlaceholder, with: resourceCategories) ) } } diff --git a/LLMonFHIR/FHIR Display/FHIRResourcesView.swift b/LLMonFHIR/FHIR Display/FHIRResourcesView.swift index f854091..bcd50ea 100644 --- a/LLMonFHIR/FHIR Display/FHIRResourcesView.swift +++ b/LLMonFHIR/FHIR Display/FHIRResourcesView.swift @@ -58,7 +58,8 @@ struct FHIRResourcesView: View { @ViewBuilder private var resourceChatView: some View { OpenAIChatView( chat: fhirMultipleResourceInterpreter.chat(resources: allResourcesArray), - title: "All FHIR Resources" + title: "All FHIR Resources", + enableFunctionCalling: true ) } diff --git a/LLMonFHIR/FHIR Display/InspectResourceView.swift b/LLMonFHIR/FHIR Display/InspectResourceView.swift index bd1d02d..2214e18 100644 --- a/LLMonFHIR/FHIR Display/InspectResourceView.swift +++ b/LLMonFHIR/FHIR Display/InspectResourceView.swift @@ -33,7 +33,8 @@ struct InspectResourceView: View { .sheet(isPresented: $showResourceChat) { OpenAIChatView( chat: fhirResourceInterpreter.chat(forResource: resource), - title: resource.displayName + title: resource.displayName, + enableFunctionCalling: false ) } .task { diff --git a/LLMonFHIR/FHIR Display/OpenAIChatView.swift b/LLMonFHIR/FHIR Display/OpenAIChatView.swift index 090caa3..8156bb9 100644 --- a/LLMonFHIR/FHIR Display/OpenAIChatView.swift +++ b/LLMonFHIR/FHIR Display/OpenAIChatView.swift @@ -15,10 +15,13 @@ import SwiftUI struct OpenAIChatView: View { @Environment(\.dismiss) private var dismiss @EnvironmentObject private var openAPIComponent: OpenAIComponent - + @EnvironmentObject private var fhirStandard: FHIR + @State private var chat: [Chat] @State private var viewState: ViewState = .idle + @State private var systemFuncMessageAdded = false + private let enableFunctionCalling: Bool private let title: String @@ -30,7 +33,7 @@ struct OpenAIChatView: View { set: { _ in } ) } - + var body: some View { NavigationStack { ChatView($chat, disableInput: disableInput) @@ -44,41 +47,38 @@ struct OpenAIChatView: View { } .viewStateAlert(state: $viewState) .onChange(of: chat) { _ in - if viewState == .idle && chat.last?.role == .user { - getAnswer() - } + getAnswer() } } } - init(chat: [Chat], title: String) { + init(chat: [Chat], title: String, enableFunctionCalling: Bool) { self._chat = State(initialValue: chat) self.title = title + self.enableFunctionCalling = enableFunctionCalling } private func getAnswer() { + guard viewState == .idle, chat.last?.role == .user else { + return + } + Task { do { viewState = .processing - let chatStreamResults = try await openAPIComponent.queryAPI(withChat: chat) - - for try await chatStreamResult in chatStreamResults { - for choice in chatStreamResult.choices { - if chat.last?.role == .assistant { - let previousChatMessage = chat.last ?? Chat(role: .assistant, content: "") - chat[chat.count - 1] = Chat( - role: .assistant, - content: (previousChatMessage.content ?? "") + (choice.delta.content ?? "") - ) - } else { - chat.append(Chat(role: .assistant, content: choice.delta.content ?? "")) - } + if enableFunctionCalling { + if systemFuncMessageAdded == false { + try await addSystemFuncMessage() + systemFuncMessageAdded = true } + try await processFunctionCalling() } + try await processChatStreamResults() + viewState = .idle } catch let error as APIErrorResponse { viewState = .error(error) @@ -87,4 +87,117 @@ struct OpenAIChatView: View { } } } + + private func addSystemFuncMessage() async throws { + let resourcesArray = await fhirStandard.resources + + var stringResourcesArray = resourcesArray.map { $0.functionCallIdentifier } + stringResourcesArray.append("N/A") + + self.chat.append(Chat(role: .system, content: String(localized: "FUNCTION_CONTEXT") + stringResourcesArray.rawValue)) + } + + private func processFunctionCalling() async throws { + let resourcesArray = await fhirStandard.resources + + var stringResourcesArray = resourcesArray.map { $0.functionCallIdentifier } + stringResourcesArray.append("N/A") + + let functionCallOutputArray = try await getFunctionCallOutputArray(stringResourcesArray) + + processFunctionCallOutputArray(functionCallOutputArray: functionCallOutputArray, resourcesArray: resourcesArray) + } + + private func getFunctionCallOutputArray(_ stringResourcesArray: [String]) async throws -> [String] { + let functions = [ + ChatFunctionDeclaration( + name: "get_resource_titles", + description: String(localized: "FUNCTION_DESCRIPTION"), + parameters: JSONSchema( + type: .object, + properties: [ + "resources": .init(type: .string, description: String(localized: "PARAMETER_DESCRIPTION"), enumValues: stringResourcesArray) + ], + required: ["resources"] + ) + ) + ] + + let chatStreamResults = try await openAPIComponent.queryAPI(withChat: chat, withFunction: functions) + + + class ChatFunctionCall { + var name: String = "" + var arguments: String = "" + var finishReason: String = "" + } + + let functionCall = ChatFunctionCall() + + for try await chatStreamResult in chatStreamResults { + for choice in chatStreamResult.choices { + if let deltaName = choice.delta.name { + functionCall.name += deltaName + } + if let deltaArguments = choice.delta.functionCall?.arguments { + functionCall.arguments += deltaArguments + } + if let finishReason = choice.finishReason { + functionCall.finishReason += finishReason + if finishReason == "get_resource_titles" { break } + } + } + } + + guard functionCall.finishReason == "function_call" else { + return [] + } + + let trimmedArguments = functionCall.arguments.trimmingCharacters(in: .whitespacesAndNewlines) + + guard let resourcesRange = trimmedArguments.range(of: "\"resources\": \"([^\"]+)\"", options: .regularExpression) else { + return [] + } + + return trimmedArguments[resourcesRange] + .replacingOccurrences(of: "\"resources\": \"", with: "") + .replacingOccurrences(of: "\"", with: "") + .components(separatedBy: ",") + } + + private func processFunctionCallOutputArray(functionCallOutputArray: [String], resourcesArray: [FHIRResource]) { + for resource in functionCallOutputArray { + guard let matchingResource = resourcesArray.first(where: { $0.functionCallIdentifier == resource }) else { + continue + } + + let functionContent = """ + Based on the function get_resource_titles you have requested the following health records: \(resource). + This is the associated JSON data for the resources which you will use to answer the users question: + \(matchingResource.jsonDescription) + + Use this health record to answer the users question ONLY IF the health record is applicable to the question. + """ + + chat.append(Chat(role: .function, content: functionContent, name: "get_resource_titles")) + } + } + + private func processChatStreamResults() async throws { + let chatStreamResults = try await openAPIComponent.queryAPI(withChat: chat) + + for try await chatStreamResult in chatStreamResults { + for choice in chatStreamResult.choices { + if chat.last?.role == .assistant { + let previousChatMessage = chat.last ?? Chat(role: .assistant, content: "") + chat[chat.count - 1] = Chat( + role: .assistant, + content: (previousChatMessage.content ?? "") + (choice.delta.content ?? "") + ) + } else { + chat.append(Chat(role: .assistant, content: choice.delta.content ?? "")) + } + } + } + } } diff --git a/LLMonFHIR/FHIR Standard/FHIRResource.swift b/LLMonFHIR/FHIR Standard/FHIRResource.swift index 2b7faf6..e6141cf 100644 --- a/LLMonFHIR/FHIR Standard/FHIRResource.swift +++ b/LLMonFHIR/FHIR Standard/FHIRResource.swift @@ -53,6 +53,10 @@ struct FHIRResource: Sendable, Identifiable, Hashable { json(withConfiguration: [.prettyPrinted, .sortedKeys, .withoutEscapingSlashes]) } + var functionCallIdentifier: String { + resourceType.filter { !$0.isWhitespace } + displayName.filter { !$0.isWhitespace } + } + private func json(withConfiguration outputFormatting: JSONEncoder.OutputFormatting) -> String { let encoder = JSONEncoder() diff --git a/LLMonFHIR/Resources/de.lproj/Localizable.strings b/LLMonFHIR/Resources/de.lproj/Localizable.strings index 682c1d8..1e703b0 100644 --- a/LLMonFHIR/Resources/de.lproj/Localizable.strings +++ b/LLMonFHIR/Resources/de.lproj/Localizable.strings @@ -126,28 +126,35 @@ Stellen Sie sich am Anfang nicht vor und beginnen Sie mit Ihrer Interpretation. "; "FHIR_MULTIPLE_RESOURCE_INTERPRETATION_PROMPT %@" = " -Sie sind die LLM on FHIR-Anwendung. -Ihre Aufgabe besteht darin, alle FHIR-Ressourcen aus den klinischen Aufzeichnungen des Benutzers zu interpretieren. +You are the LLM on FHIR application. +Your task is to interpret FHIR resources from the user's clinical records. +Throughout the conversation with the user, use the get_resource_titles function to obtain the FHIR health resources neccesary to properly answer the users question. For example, if the user asks about their allergies, you must use get_resource_titles to output the FHIR resource titles for allergy records so you can then use them to answer the question. The output of get_resource_titles has to be the name of a resource or resources with the exact same title as in the list provided. -Interpretieren Sie alle Ressourcen, indem Sie ihre Daten erklären, die für die Gesundheit des Benutzers relevant sind. -Erklären Sie den relevanten medizinischen Kontext in einer für einen Nicht-Mediziner verständlichen Sprache. -Sie sollten sachliche und präzise Informationen in einer kompakten Zusammenfassung in kurzen Antworten geben. +Answer the users question, the health record provided is not always related. The end goal is to answer the users question in the best way possible. -Die folgende JSON-Repräsentation definiert die FHIR-Ressourcen, die Sie interpretieren sollten: -%@ +Interpret the resources by explaining its data relevant to the user's health. +Explain the relevant medical context in a language understandable by a user who is not a medical professional. +You should provide factual and precise information in a compact summary in short responses. -Informieren Sie den Benutzer, dass er Fragen zu seinen Gesundheitsaufzeichnungen stellen kann, und erstellen Sie dann eine kurze Liste der Hauptkategorien der Gesundheitsaufzeichnungen des Benutzers, auf die Sie Zugriff haben. +Tell the user that they can ask any question about their health records and then create a short summary of the main categories of health records of the user which you have access to. These are the resource titles: +%@ -Geben Sie dem Benutzer sofort eine Interpretation, um das Gespräch zu beginnen. -Die erste Interpretation sollte eine kurze und einfache Zusammenfassung mit folgenden Spezifikationen sein: -1. Gesamtzusammenfassung aller Gesundheitsaufzeichnungen -2. Lesestufe Mittelschule -3. Schließen Sie mit einer Frage ab, in der der Benutzer gefragt wird, ob er Fragen hat. Stellen Sie sicher, dass diese Frage nicht allgemein, sondern spezifisch für ihre Gesundheitsaufzeichnungen ist. -Stellen Sie sich am Anfang nicht vor und beginnen Sie mit Ihrer Interpretation. -Stellen Sie sicher, dass Ihre Antwort in derselben Sprache erfolgt, in der der Benutzer Ihnen schreibt. -Die Zeitform sollte in der Gegenwart sein. +Immediately return a short summary of the users health records to start the conversation. +The initial summary should be a short and simple summary with the following specifications: +1. Short summary of the health records categories +2. 5th grade reading level +3. End with a question asking user if they have any questions. Make sure that this question is not generic but specific to their health records. +Do not introduce yourself at the beginning, and start with your interpretation. +Make sure your response is in the same language the user writes to you in. +The tense should be present. "; +"FUNCTION_DESCRIPTION" = "Call this function to determine the relevant FHIR health record titles based on the user's question. The titles must have the exact name as in the enumValues array given. A question can build upon a previous question and does not need to explicilty state a health record. The function will decide which health record titles are applicable to the question and it can return multiple titles or N/A if no category is suitable. Only output a record title/titles if it is directly applicable to the question otherwise output N/A. Always try to output the least amount of resource titles to be sent to the model to prevent exceeding the token limit."; + +"PARAMETER_DESCRIPTION" = "Provide a comma-separated list of all the FHIR health record titles with the EXACT SAME NAME AS GIVEN IN THE LIST that are applicable to answer the user's questions. These titles have to be the SAME NAME AS GIVEN IN THE ARRAY provided. If multiple titles apply, separate each title by a comma and space (e.g for multiple medications). Try to provide all the required titles to allow yourself to fully answer the question in a comprehensive manner. If the question doesn't fit into any category, output N/A. For example, if a user asks: 'Tell me more about my medications,' then output all titles associated with medications. A question can build upon a previous question and does not need to be explicit. e.g. if a user says prescribe, this is associated with medication. Do not exceed token limit with outputted titles."; + +"FUNCTION_CONTEXT" = "Use the function call 'get_resource_titles' and provide a comma-separated list resource titles directly applicable to the question. The function will determine which FHIR health record titles are relevant to the user's question using this array: "; + // MARK: - Settings "SETTINGS_TITLE" = "Einstellungen"; diff --git a/LLMonFHIR/Resources/en.lproj/Localizable.strings b/LLMonFHIR/Resources/en.lproj/Localizable.strings index cd926c7..5e35dd5 100644 --- a/LLMonFHIR/Resources/en.lproj/Localizable.strings +++ b/LLMonFHIR/Resources/en.lproj/Localizable.strings @@ -128,27 +128,34 @@ Do not introduce yourself at the beginning, and start with your interpretation. "FHIR_MULTIPLE_RESOURCE_INTERPRETATION_PROMPT %@" = " You are the LLM on FHIR application. -Your task is to interpret all of the FHIR resources from the user's clinical records. +Your task is to interpret FHIR resources from the user's clinical records. +Throughout the conversation with the user, use the get_resource_titles function to obtain the FHIR health resources neccesary to properly answer the users question. For example, if the user asks about their allergies, you must use get_resource_titles to output the FHIR resource titles for allergy records so you can then use them to answer the question. The output of get_resource_titles has to be the name of a resource or resources with the exact same title as in the list provided. -Interpret all the resources by explaining its data relevant to the user's health. +Answer the users question, the health record provided is not always related. The end goal is to answer the users question in the best way possible. + +Interpret the resources by explaining its data relevant to the user's health. Explain the relevant medical context in a language understandable by a user who is not a medical professional. You should provide factual and precise information in a compact summary in short responses. -The following JSON representation defines the FHIR resources that you should interpret: +Tell the user that they can ask any question about their health records and then create a short summary of the main categories of health records of the user which you have access to. These are the resource titles: %@ -Tell the user that they can ask any question about their health records and then create a short list of the main categories of health records of the user which you have access to. - -Immediately return an interpretation to the user, starting the conversation. -The initial interpretation should be a short and simple summary with the following specifications: -1. Overall summary of all health records -2. Middle school reading level +Immediately return a short summary of the users health records to start the conversation. +The initial summary should be a short and simple summary with the following specifications: +1. Short summary of the health records categories +2. 5th grade reading level 3. End with a question asking user if they have any questions. Make sure that this question is not generic but specific to their health records. Do not introduce yourself at the beginning, and start with your interpretation. Make sure your response is in the same language the user writes to you in. The tense should be present. "; +"FUNCTION_DESCRIPTION" = "Call this function to determine the relevant FHIR health record titles based on the user's question. The titles must have the exact name as in the enumValues array given. A question can build upon a previous question and does not need to explicilty state a health record. The function will decide which health record titles are applicable to the question and it can return multiple titles or N/A if no category is suitable. Only output a record title/titles if it is directly applicable to the question otherwise output N/A. Always try to output the least amount of resource titles to be sent to the model to prevent exceeding the token limit."; + +"PARAMETER_DESCRIPTION" = "Provide a comma-separated list of all the FHIR health record titles with the EXACT SAME NAME AS GIVEN IN THE LIST that are applicable to answer the user's questions. These titles have to be the SAME NAME AS GIVEN IN THE ARRAY provided. If multiple titles apply, separate each title by a comma and space (e.g for multiple medications). Try to provide all the required titles to allow yourself to fully answer the question in a comprehensive manner. If the question doesn't fit into any category, output N/A. For example, if a user asks: 'Tell me more about my medications,' then output all titles associated with medications. A question can build upon a previous question and does not need to be explicit. e.g. if a user says prescribe, this is associated with medication. Do not exceed token limit with outputted titles."; + +"FUNCTION_CONTEXT" = "Use the function call 'get_resource_titles' and provide a comma-separated list resource titles directly applicable to the question. The function will determine which FHIR health record titles are relevant to the user's question using this array: "; + // MARK: - Settings "SETTINGS_TITLE" = "Settings"; diff --git a/LLMonFHIR/Resources/es.lproj/Localizable.strings b/LLMonFHIR/Resources/es.lproj/Localizable.strings index 90c6922..2e1d339 100644 --- a/LLMonFHIR/Resources/es.lproj/Localizable.strings +++ b/LLMonFHIR/Resources/es.lproj/Localizable.strings @@ -126,28 +126,35 @@ No se presente al principio, y comience con su interpretación. "; "FHIR_MULTIPLE_RESOURCE_INTERPRETATION_PROMPT %@" = " -Usted es la aplicación LLM en FHIR. -Su tarea es interpretar todos los recursos FHIR de los registros clínicos del usuario. +You are the LLM on FHIR application. +Your task is to interpret FHIR resources from the user's clinical records. +Throughout the conversation with the user, use the get_resource_titles function to obtain the FHIR health resources neccesary to properly answer the users question. For example, if the user asks about their allergies, you must use get_resource_titles to output the FHIR resource titles for allergy records so you can then use them to answer the question. The output of get_resource_titles has to be the name of a resource or resources with the exact same title as in the list provided. -Interprete todos los recursos explicando sus datos relevantes para la salud del usuario. -Explique el contexto médico relevante en un lenguaje comprensible para un usuario que no es un profesional médico. -Debe proporcionar información factual y precisa en un resumen compacto en respuestas cortas. +Answer the users question, the health record provided is not always related. The end goal is to answer the users question in the best way possible. -La siguiente representación JSON define todos los recursos FHIR que debe interpretar: -%@ +Interpret the resources by explaining its data relevant to the user's health. +Explain the relevant medical context in a language understandable by a user who is not a medical professional. +You should provide factual and precise information in a compact summary in short responses. -Informar al usuario que puede hacer preguntas sobre sus registros de salud y luego crear una breve lista de las principales categorías de registros de salud del usuario a las que tiene acceso. +Tell the user that they can ask any question about their health records and then create a short summary of the main categories of health records of the user which you have access to. These are the resource titles: +%@ -Devuelva una interpretación al usuario de inmediato para comenzar la conversación. -La interpretación inicial debe ser un resumen breve y simple con las siguientes especificaciones: -1. Resumen general de todos los registros de salud -2. Nivel de lectura de la escuela secundaria -3. Terminar con una pregunta que le pregunte al usuario si tiene alguna pregunta. Asegúrese de que esta pregunta no sea genérica, sino específica para sus registros de salud. -No se presente al principio, y comience con su interpretación. -Asegúrese de que su respuesta esté en el mismo idioma en que el usuario le escribe. -El tiempo verbal debe ser presente. +Immediately return a short summary of the users health records to start the conversation. +The initial summary should be a short and simple summary with the following specifications: +1. Short summary of the health records categories +2. 5th grade reading level +3. End with a question asking user if they have any questions. Make sure that this question is not generic but specific to their health records. +Do not introduce yourself at the beginning, and start with your interpretation. +Make sure your response is in the same language the user writes to you in. +The tense should be present. "; +"FUNCTION_DESCRIPTION" = "Call this function to determine the relevant FHIR health record titles based on the user's question. The titles must have the exact name as in the enumValues array given. A question can build upon a previous question and does not need to explicilty state a health record. The function will decide which health record titles are applicable to the question and it can return multiple titles or N/A if no category is suitable. Only output a record title/titles if it is directly applicable to the question otherwise output N/A. Always try to output the least amount of resource titles to be sent to the model to prevent exceeding the token limit."; + +"PARAMETER_DESCRIPTION" = "Provide a comma-separated list of all the FHIR health record titles with the EXACT SAME NAME AS GIVEN IN THE LIST that are applicable to answer the user's questions. These titles have to be the SAME NAME AS GIVEN IN THE ARRAY provided. If multiple titles apply, separate each title by a comma and space (e.g for multiple medications). Try to provide all the required titles to allow yourself to fully answer the question in a comprehensive manner. If the question doesn't fit into any category, output N/A. For example, if a user asks: 'Tell me more about my medications,' then output all titles associated with medications. A question can build upon a previous question and does not need to be explicit. e.g. if a user says prescribe, this is associated with medication. Do not exceed token limit with outputted titles."; + +"FUNCTION_CONTEXT" = "Use the function call 'get_resource_titles' and provide a comma-separated list resource titles directly applicable to the question. The function will determine which FHIR health record titles are relevant to the user's question using this array: "; + // MARK: - Settings "SETTINGS_TITLE" = "Configuración"; diff --git a/LLMonFHIR/Resources/fr.lproj/Localizable.strings b/LLMonFHIR/Resources/fr.lproj/Localizable.strings index 18e180d..dc9d2a3 100644 --- a/LLMonFHIR/Resources/fr.lproj/Localizable.strings +++ b/LLMonFHIR/Resources/fr.lproj/Localizable.strings @@ -126,28 +126,35 @@ Ne vous présentez pas au début, et commencez par votre interprétation. "; "FHIR_MULTIPLE_RESOURCE_INTERPRETATION_PROMPT %@" = " -Vous êtes l'application LLM sur FHIR. -Votre tâche consiste à interpréter toutes les ressources FHIR à partir des dossiers cliniques de l'utilisateur. +You are the LLM on FHIR application. +Your task is to interpret FHIR resources from the user's clinical records. +Throughout the conversation with the user, use the get_resource_titles function to obtain the FHIR health resources neccesary to properly answer the users question. For example, if the user asks about their allergies, you must use get_resource_titles to output the FHIR resource titles for allergy records so you can then use them to answer the question. The output of get_resource_titles has to be the name of a resource or resources with the exact same title as in the list provided. -Interprétez toutes les ressources en expliquant leurs données pertinentes pour la santé de l'utilisateur. -Expliquez le contexte médical pertinent dans un langage compréhensible pour un utilisateur qui n'est pas un professionnel de la santé. -Vous devez fournir des informations factuelles et précises dans un résumé compact en réponses courtes. +Answer the users question, the health record provided is not always related. The end goal is to answer the users question in the best way possible. -La représentation JSON suivante définit toutes les ressources FHIR que vous devez interpréter : -%@ +Interpret the resources by explaining its data relevant to the user's health. +Explain the relevant medical context in a language understandable by a user who is not a medical professional. +You should provide factual and precise information in a compact summary in short responses. -Informez l'utilisateur qu'il peut poser des questions sur ses dossiers de santé, puis créez une courte liste des principales catégories de dossiers de santé de l'utilisateur auxquelles vous avez accès. +Tell the user that they can ask any question about their health records and then create a short summary of the main categories of health records of the user which you have access to. These are the resource titles: +%@ -Retournez une interprétation à l'utilisateur immédiatement pour commencer la conversation. -L'interprétation initiale doit être un bref résumé simple avec les spécifications suivantes : -1. Résumé général de tous les dossiers de santé -2. Niveau de lecture de l'école secondaire -3. Terminez avec une question demandant à l'utilisateur s'il a des questions. Assurez-vous que cette question n'est pas générique, mais spécifique à ses dossiers de santé. -Ne vous présentez pas au début, et commencez par votre interprétation. -Assurez-vous que votre réponse est dans la même langue que celle de l'utilisateur. -Le temps verbal doit être au présent. +Immediately return a short summary of the users health records to start the conversation. +The initial summary should be a short and simple summary with the following specifications: +1. Short summary of the health records categories +2. 5th grade reading level +3. End with a question asking user if they have any questions. Make sure that this question is not generic but specific to their health records. +Do not introduce yourself at the beginning, and start with your interpretation. +Make sure your response is in the same language the user writes to you in. +The tense should be present. "; +"FUNCTION_DESCRIPTION" = "Call this function to determine the relevant FHIR health record titles based on the user's question. The titles must have the exact name as in the enumValues array given. A question can build upon a previous question and does not need to explicilty state a health record. The function will decide which health record titles are applicable to the question and it can return multiple titles or N/A if no category is suitable. Only output a record title/titles if it is directly applicable to the question otherwise output N/A. Always try to output the least amount of resource titles to be sent to the model to prevent exceeding the token limit."; + +"PARAMETER_DESCRIPTION" = "Provide a comma-separated list of all the FHIR health record titles with the EXACT SAME NAME AS GIVEN IN THE LIST that are applicable to answer the user's questions. These titles have to be the SAME NAME AS GIVEN IN THE ARRAY provided. If multiple titles apply, separate each title by a comma and space (e.g for multiple medications). Try to provide all the required titles to allow yourself to fully answer the question in a comprehensive manner. If the question doesn't fit into any category, output N/A. For example, if a user asks: 'Tell me more about my medications,' then output all titles associated with medications. A question can build upon a previous question and does not need to be explicit. e.g. if a user says prescribe, this is associated with medication. Do not exceed token limit with outputted titles."; + +"FUNCTION_CONTEXT" = "Use the function call 'get_resource_titles' and provide a comma-separated list resource titles directly applicable to the question. The function will determine which FHIR health record titles are relevant to the user's question using this array: "; + // MARK: - Settings "SETTINGS_TITLE" = "Paramètres"; diff --git a/LLMonFHIR/Resources/zh-Hans.lproj/Localizable.strings b/LLMonFHIR/Resources/zh-Hans.lproj/Localizable.strings index 11bab7d..e382de8 100644 --- a/LLMonFHIR/Resources/zh-Hans.lproj/Localizable.strings +++ b/LLMonFHIR/Resources/zh-Hans.lproj/Localizable.strings @@ -126,28 +126,35 @@ Chat with the user in %@"; "; "FHIR_MULTIPLE_RESOURCE_INTERPRETATION_PROMPT %@" = " -您是LLM on FHIR应用程序。 -您的任务是解释用户临床记录中的所有FHIR资源。 +You are the LLM on FHIR application. +Your task is to interpret FHIR resources from the user's clinical records. +Throughout the conversation with the user, use the get_resource_titles function to obtain the FHIR health resources neccesary to properly answer the users question. For example, if the user asks about their allergies, you must use get_resource_titles to output the FHIR resource titles for allergy records so you can then use them to answer the question. The output of get_resource_titles has to be the name of a resource or resources with the exact same title as in the list provided. -通过解释与用户健康相关的数据来解释所有资源。 -用非医学专业人士能理解的语言解释相关的医学背景。 -您应该在简短的回答中提供事实和准确的信息，以紧凑的摘要形式。 +Answer the users question, the health record provided is not always related. The end goal is to answer the users question in the best way possible. -以下JSON表示定义了您应该解释的所有FHIR资源： -%@ +Interpret the resources by explaining its data relevant to the user's health. +Explain the relevant medical context in a language understandable by a user who is not a medical professional. +You should provide factual and precise information in a compact summary in short responses. -告诉用户他们可以就他们的健康记录询问任何问题，然后简要列出您可以访问的用户主要健康记录的主要类别。 +Tell the user that they can ask any question about their health records and then create a short summary of the main categories of health records of the user which you have access to. These are the resource titles: +%@ -立即向用户返回解释，开始对话。 -首次解释应该是简短且简单的总结，具体规格如下： -1. 所有健康记录的总体摘要 -2. 中学阅读水平 -3. 结尾时询问用户是否有任何问题。确保这个问题不是泛泛而谈，而是特定于他们的健康记录。 -在开始时不要介绍自己，直接开始解释。 -确保您的回应用与用户写给您的语言相同。 -动词时态应为现在时。 +Immediately return a short summary of the users health records to start the conversation. +The initial summary should be a short and simple summary with the following specifications: +1. Short summary of the health records categories +2. 5th grade reading level +3. End with a question asking user if they have any questions. Make sure that this question is not generic but specific to their health records. +Do not introduce yourself at the beginning, and start with your interpretation. +Make sure your response is in the same language the user writes to you in. +The tense should be present. "; +"FUNCTION_DESCRIPTION" = "Call this function to determine the relevant FHIR health record titles based on the user's question. The titles must have the exact name as in the enumValues array given. A question can build upon a previous question and does not need to explicilty state a health record. The function will decide which health record titles are applicable to the question and it can return multiple titles or N/A if no category is suitable. Only output a record title/titles if it is directly applicable to the question otherwise output N/A. Always try to output the least amount of resource titles to be sent to the model to prevent exceeding the token limit."; + +"PARAMETER_DESCRIPTION" = "Provide a comma-separated list of all the FHIR health record titles with the EXACT SAME NAME AS GIVEN IN THE LIST that are applicable to answer the user's questions. These titles have to be the SAME NAME AS GIVEN IN THE ARRAY provided. If multiple titles apply, separate each title by a comma and space (e.g for multiple medications). Try to provide all the required titles to allow yourself to fully answer the question in a comprehensive manner. If the question doesn't fit into any category, output N/A. For example, if a user asks: 'Tell me more about my medications,' then output all titles associated with medications. A question can build upon a previous question and does not need to be explicit. e.g. if a user says prescribe, this is associated with medication. Do not exceed token limit with outputted titles."; + +"FUNCTION_CONTEXT" = "Use the function call 'get_resource_titles' and provide a comma-separated list resource titles directly applicable to the question. The function will determine which FHIR health record titles are relevant to the user's question using this array: "; + // MARK: - Settings "SETTINGS_TITLE" = "设置";