feat: Move Inference Config to Prompt Settings

Author: kieran-wilkinson-4

packages/cdk/resources/BedrockPromptResources.ts

@@ -43,8 +43,8 @@ export class BedrockPromptResources extends Construct {
       messages: [props.settings.userPrompt]
     })
 
-    ragResponsePromptVariant["inferenceConfiguration"] = {
-      "text": props.settings.inferenceConfig
+    ragResponsePromptVariant.inferenceConfiguration = {
+      text: props.settings.inferenceConfig
     }
 
     const ragPrompt = new Prompt(this, "ragResponsePrompt", {

packages/cdk/resources/BedrockPromptSettings.ts

@@ -1,16 +1,10 @@
 import * as fs from "fs"
 import {ChatMessage} from "@cdklabs/generative-ai-cdk-constructs/lib/cdk-lib/bedrock"
 import {Construct} from "constructs"
+import {CfnPrompt} from "aws-cdk-lib/aws-bedrock"
 
 export type BedrockPromptSettingsType = "system" | "user" | "reformulation"
 
-export interface BedrockPromptInferenceConfig {
-  temperature: number,
-  topP: number,
-  maxTokens: number,
-  stopSequences: Array<string>
-}
-
 /** BedrockPromptSettings is responsible for loading and providing
  * the system, user, and reformulation prompts along with their
  * inference configurations.
@@ -19,7 +13,7 @@ export class BedrockPromptSettings extends Construct {
   public readonly systemPrompt: ChatMessage
   public readonly userPrompt: ChatMessage
   public readonly reformulationPrompt: ChatMessage
-  public readonly inferenceConfig: BedrockPromptInferenceConfig
+  public readonly inferenceConfig: CfnPrompt.PromptModelInferenceConfigurationProperty
 
   /**
    * @param scope The Construct scope
@@ -38,14 +32,10 @@ export class BedrockPromptSettings extends Construct {
     const reformulationPrompt = this.getTypedPrompt("reformulation")
     this.reformulationPrompt = ChatMessage.user(reformulationPrompt.text)
 
-    const temperature = this.node.tryGetContext("ragTemperature")
-    const maxTokens = this.node.tryGetContext("ragMaxTokens")
-    const topP = this.node.tryGetContext("ragTopP")
-
     this.inferenceConfig = {
-      temperature: parseInt(temperature, 10),
-      topP: parseInt(topP, 10),
-      maxTokens: parseInt(maxTokens, 10),
+      temperature: 0,
+      topP: 1,
+      maxTokens: 512,
       stopSequences: [
         "Human:"
       ]

packages/cdk/resources/Functions.ts

@@ -11,10 +11,6 @@ const RAG_MODEL_ID = "anthropic.claude-3-sonnet-20240229-v1:0"
 const QUERY_REFORMULATION_MODEL_ID = "anthropic.claude-3-haiku-20240307-v1:0"
 const BEDROCK_KB_DATA_SOURCE = "eps-assist-kb-ds"
 const LAMBDA_MEMORY_SIZE = "265"
-// Claude RAG inference parameters
-const RAG_TEMPERATURE = "0"
-const RAG_MAX_TOKENS = "512"
-const RAG_TOP_P = "1"
 
 export interface FunctionsProps {
   readonly stackName: string
@@ -75,10 +71,7 @@ export class Functions extends Construct {
         "QUERY_REFORMULATION_PROMPT_NAME": props.reformulationPromptName,
         "RAG_RESPONSE_PROMPT_NAME": props.ragResponsePromptName,
         "QUERY_REFORMULATION_PROMPT_VERSION": props.reformulationPromptVersion,
-        "RAG_RESPONSE_PROMPT_VERSION": props.ragResponsePromptVersion,
-        "RAG_TEMPERATURE": RAG_TEMPERATURE,
-        "RAG_MAX_TOKENS": RAG_MAX_TOKENS,
-        "RAG_TOP_P": RAG_TOP_P
+        "RAG_RESPONSE_PROMPT_VERSION": props.ragResponsePromptVersion
       }
     })