Update: [AEA-6157] - Exceed Context Window Issue

packages/cdk/nagSuppressions.ts

@@ -80,6 +80,22 @@ export const nagSuppressions = (stack: Stack, account: string) => {
     ]
   )
 
+  // Suppress unauthenticated API route warnings
+  safeAddNagSuppression(
+    stack,
+    "/EpsAssistMeStack/Apis/EpsAssistApiGateway/ApiGateway/Default/slack/commands/POST/Resource",
+    [
+      {
+        id: "AwsSolutions-APIG4",
+        reason: "Slack command endpoint is intentionally unauthenticated."
+      },
+      {
+        id: "AwsSolutions-COG4",
+        reason: "Cognito not required for this public endpoint."
+      }
+    ]
+  )
+
   // Suppress missing WAF on API stage for Apis construct
   safeAddNagSuppression(
     stack,

packages/cdk/prompts/reformulationPrompt.txt

@@ -1,5 +1,40 @@
-Return the user query exactly as provided without any modifications, changes, or reformulations.
-Do not alter, rephrase, or modify the input in any way.
-Simply return: {{user_query}}
+<|begin_of_text|><|start_header_id|>system<|end_header_id|>
+You are an expert RAG query and context optimizer. Your task is to analyze verbose user queries and raw search context, stripping away all conversational filler to output a concise, impactful summary.
 
-User Query: {{user_query}}
+You must:
+1. Extract the core objective into a single, direct question.
+2. Capture individual questions and their specific needs. 
+3. Isolate critical variables, specific states, and constraints required to solve the problem.
+4. Enhance the question(s) with relevant terminology from the search results
+
+Output your response strictly using the following XML structure:
+<optimized_query> (The short, direct question)
+<key_variables> (Bullet points of critical states, statuses, or constraints)
+<|eot_id|>
+
+<|start_header_id|>user<|end_header_id|>
+### User Query
+Hi, I need some help figuring out the PTO rules for one of my team members. They started as part-time 6 months ago, but they just transitioned to full-time last week (let's say exactly 7 days ago). They currently have 2 days of PTO saved up from their part-time stint. They want to take next week off entirely, which would require 5 days of PTO. Can they do this, effectively going to a -3 balance, since they are full-time now?
+<|eot_id|>
+
+<|start_header_id|>assistant<|end_header_id|>
+<optimized_query>
+Can a recently transitioned full-time employee with 6 months total tenure and 2 accrued PTO days take 5 days off, resulting in a -3 PTO balance?
+</optimized_query>
+<key_variables>
+- Current Status: Full-time (transitioned 7 days ago)
+- Total Tenure: 6 months
+- Current PTO Balance: 2 days
+- Requested PTO: 5 days (resulting in -3 balance)
+</key_variables>
+<|eot_id|>
+
+<|start_header_id|>user<|end_header_id|>
+### Search Context
+$search_results$
+
+### User Query
+{{user_query}}
+<|eot_id|>
+
+<|start_header_id|>assistant<|end_header_id|>

packages/cdk/prompts/systemPrompt.txt

@@ -1,24 +1,32 @@
-# 1. Persona & Logic
-You are an AI assistant for onboarding guidance. Follow these strict rules:
-- **Strict Evidence:** If the answer is missing, do not infer or use external knowledge. 
-- **Grounding:** NEVER use your own internal training data, online resources, or prior knowledge.
-- **Decomposition:** Split multi-part queries into numbered sub-questions (Q1, Q2).
+You are a technical assistant specialized in onboarding guidance. 
+Your primary goal is to
 
-# 2. Output Structure
-**Summary**
-2-3 sentences maximum.
+STYLE & FORMATTING RULES:
+- Do NOT refer to the search results by number or name in the body of the text.
+- Do NOT add a "Citations" section at the end of the response.
+- Do NOT reference how the information was found (e.g., "...the provided search results")
+- Do NOT state what the data is related to (e.g., "The search results are related to NHS API and FHIR...")
+- Text should prioritise readability.
+- Links should use Markdown text, e.g., <url|link text>.
+- Use `Inline Code` for system names, field names, or technical terms (e.g., `HL7 FHIR`).
 
- **Answer**
- Prioritize detail and specification, focus on the information direct at the question.
+RULES:
+-  Answer questions using ONLY the provided search results. 
+- Do not assume any information, all information must be grounded in data.
 
-# 3. Styling Rules (`mrkdwn`)
-Use British English.
-- **Bold (`*`):** Headings, Subheadings, Source Names, and important information/ exceptions (e.g. `*NHS England*`).
-- **Italic (`_`):** Citations and Titles (e.g. `_Guidance v1_`).
-- **Blockquote (`>`):** Quotes (>1 sentence) and Tech Specs/Examples  (e.g. `HL7 FHIR`).
-- **Links:** `[text](link)`.
+STEPS:
+1. Extract key information from the knowledge base
+2. Generate an answer, capturing the core question the user is asking.
+3. Answer, directly, any individual or sub-questions the user has provided.
+4. You must create a very short summary encapsulating the response and have it precede all other answers.
 
-# 4. Format Rules
-- NEVER use in-line references or citations (e.g., do not write "(search result 1)" or "[1]").
-- Do NOT refer to the search results by number or name in the body of the text.
-- Do NOT add a "Citations" section at the end of the response.wer, details from the knowledge base.
+EXAMPLE:
+<example_interaction>
+*Summary*
+This is a short, fast answer so the user doesn't _have_ to read the long answer.
+
+*Answer*
+This is a direct answer to the question, or questions, provided. It breaks down individual questions. There is no reference to the text here (for example, you don't see "from source 1") but instead treats this information as if it was public knowledge. However, if there is a source, it does provide that source [as a hyperlink](hyperlink) to the website it can be found.
+
+There is multiple paragraphs, with blank lines between, to make it easier to read, as readability is a requirement.
+</example_interaction>

packages/cdk/prompts/userPrompt.txt

@@ -1,4 +1,3 @@
-
 <search_results>$search_results$<search_results>
 
 <user_query>{{user_query}}<user_query>

packages/cdk/resources/Apis.ts

@@ -27,6 +27,7 @@ export class Apis extends Construct {
       forwardCsocLogs: props.forwardCsocLogs,
       csocApiGatewayDestination: props.csocApiGatewayDestination
     })
+
     // Create /slack resource path
     const slackResource = apiGateway.api.root.addResource("slack")
 
@@ -41,6 +42,17 @@ export class Apis extends Construct {
       lambdaFunction: props.functions.slackBot
     })
 
+    // Create the '/slack/commands' POST endpoint for Slack Events API
+    // This endpoint will handle slash commands, such as /test
+    // eslint-disable-next-line @typescript-eslint/no-unused-vars
+    const slackCommandsEndpoint = new LambdaEndpoint(this, "SlackCommandsEndpoint", {
+      parentResource: slackResource,
+      resourceName: "commands",
+      method: HttpMethod.POST,
+      restApiGatewayRole: apiGateway.role,
+      lambdaFunction: props.functions.slackBot
+    })
+
     this.apis = {
       api: apiGateway
     }

packages/cdk/resources/BedrockPromptResources.ts

@@ -1,66 +1,85 @@
 import {Construct} from "constructs"
+import * as crypto from "crypto"
 import {
   BedrockFoundationModel,
+  ChatMessage,
   Prompt,
   PromptVariant
 } from "@cdklabs/generative-ai-cdk-constructs/lib/cdk-lib/bedrock"
 import {BedrockPromptSettings} from "./BedrockPromptSettings"
+import {CfnPrompt} from "aws-cdk-lib/aws-bedrock"
 
 export interface BedrockPromptResourcesProps {
   readonly stackName: string
   readonly settings: BedrockPromptSettings
 }
 
 export class BedrockPromptResources extends Construct {
-  public readonly queryReformulationPrompt: Prompt
+  public readonly reformulationPrompt: Prompt
   public readonly ragResponsePrompt: Prompt
-  public readonly ragModelId: string
-  public readonly queryReformulationModelId: string
+  public readonly modelId: string
 
   constructor(scope: Construct, id: string, props: BedrockPromptResourcesProps) {
     super(scope, id)
 
-    const ragModel = new BedrockFoundationModel("meta.llama3-70b-instruct-v1:0")
-    const reformulationModel = BedrockFoundationModel.AMAZON_NOVA_LITE_V1
+    const aiModel = new BedrockFoundationModel("meta.llama3-70b-instruct-v1:0")
 
-    const queryReformulationPromptVariant = PromptVariant.text({
-      variantName: "default",
-      model: reformulationModel,
-      promptVariables: ["topic"],
-      promptText: props.settings.reformulationPrompt.text
-    })
+    // Create Prompts
+    this.reformulationPrompt = this.createPrompt(
+      "ReformulationPrompt",
+      `${props.stackName}-reformulation`,
+      "Prompt for reformulation queries to improve RAG inference",
+      aiModel,
+      "",
+      [props.settings.reformulationPrompt],
+      props.settings.reformulationInferenceConfig
+    )
 
-    const queryReformulationPrompt = new Prompt(this, "QueryReformulationPrompt", {
-      promptName: `${props.stackName}-queryReformulation`,
-      description: "Prompt for reformulating user queries to improve RAG retrieval",
-      defaultVariant: queryReformulationPromptVariant,
-      variants: [queryReformulationPromptVariant]
-    })
+    this.ragResponsePrompt = this.createPrompt(
+      "RagResponsePrompt",
+      `${props.stackName}-ragResponse`,
+      "Prompt for generating RAG responses with knowledge base context and system instructions",
+      aiModel,
+      props.settings.systemPrompt.text,
+      [props.settings.userPrompt],
+      props.settings.ragInferenceConfig
+    )
+
+    this.modelId = aiModel.modelId
+  }
 
-    const ragResponsePromptVariant = PromptVariant.chat({
+  private createPrompt(
+    id: string,
+    promptName: string,
+    description: string,
+    model: BedrockFoundationModel,
+    systemPromptText: string,
+    messages: [ChatMessage],
+    inferenceConfig: CfnPrompt.PromptModelInferenceConfigurationProperty
+  ): Prompt {
+
+    const variant = PromptVariant.chat({
       variantName: "default",
-      model: ragModel,
-      promptVariables: ["query", "search_results"],
-      system: props.settings.systemPrompt.text,
-      messages: [props.settings.userPrompt]
+      model: model,
+      promptVariables: ["prompt", "search_results"],
+      system: systemPromptText,
+      messages: messages
     })
 
-    ragResponsePromptVariant.inferenceConfiguration = {
-      text: props.settings.inferenceConfig
+    variant.inferenceConfiguration = {
+      text: inferenceConfig
     }
 
-    const ragPrompt = new Prompt(this, "ragResponsePrompt", {
-      promptName: `${props.stackName}-ragResponse`,
-      description: "Prompt for generating RAG responses with knowledge base context and system instructions",
-      defaultVariant: ragResponsePromptVariant,
-      variants: [ragResponsePromptVariant]
-    })
-
-    // expose model IDs for use in Lambda environment variables
-    this.ragModelId = ragModel.modelId
-    this.queryReformulationModelId = reformulationModel.modelId
+    const hash = crypto.createHash("md5")
+      .update(JSON.stringify(variant))
+      .digest("hex")
+      .substring(0, 6)
 
-    this.queryReformulationPrompt = queryReformulationPrompt
-    this.ragResponsePrompt = ragPrompt
+    return new Prompt(this, id, {
+      promptName: `${promptName}-${hash}`,
+      description,
+      defaultVariant: variant,
+      variants: [variant]
+    })
   }
 }

packages/cdk/resources/BedrockPromptSettings.ts

@@ -3,7 +3,7 @@ import {ChatMessage} from "@cdklabs/generative-ai-cdk-constructs/lib/cdk-lib/bed
 import {Construct} from "constructs"
 import {CfnPrompt} from "aws-cdk-lib/aws-bedrock"
 
-export type BedrockPromptSettingsType = "system" | "user" | "reformulation"
+export type BedrockPromptSettingsType = "system" | "reformulation" | "user"
 
 /** BedrockPromptSettings is responsible for loading and providing
  * the system, user, and reformulation prompts along with their
@@ -13,7 +13,8 @@ export class BedrockPromptSettings extends Construct {
   public readonly systemPrompt: ChatMessage
   public readonly userPrompt: ChatMessage
   public readonly reformulationPrompt: ChatMessage
-  public readonly inferenceConfig: CfnPrompt.PromptModelInferenceConfigurationProperty
+  public readonly ragInferenceConfig: CfnPrompt.PromptModelInferenceConfigurationProperty
+  public readonly reformulationInferenceConfig: CfnPrompt.PromptModelInferenceConfigurationProperty
 
   /**
    * @param scope The Construct scope
@@ -30,16 +31,19 @@ export class BedrockPromptSettings extends Construct {
     this.userPrompt = ChatMessage.user(userPromptData.text)
 
     const reformulationPrompt = this.getTypedPrompt("reformulation")
-    this.reformulationPrompt = ChatMessage.user(reformulationPrompt.text)
+    this.reformulationPrompt = ChatMessage.assistant(reformulationPrompt.text)
 
-    this.inferenceConfig = {
+    const defaultInferenceConfig = {
       temperature: 0,
       topP: 0.3,
-      maxTokens: 1024,
+      maxTokens: 512,
       stopSequences: [
         "Human:"
       ]
     }
+
+    this.ragInferenceConfig = defaultInferenceConfig
+    this.reformulationInferenceConfig = defaultInferenceConfig
   }
 
   /** Get the latest prompt text from files in the specified directory.

packages/cdk/resources/Functions.ts

@@ -35,7 +35,7 @@ export interface FunctionsProps {
   readonly isPullRequest: boolean
   readonly mainSlackBotLambdaExecutionRoleArn : string
   readonly ragModelId: string
-  readonly queryReformulationModelId: string
+  readonly reformulationModelId: string
   readonly notifyS3UploadFunctionPolicy: ManagedPolicy
   readonly docsBucketName: string
 }
@@ -61,17 +61,17 @@ export class Functions extends Construct {
       dependencyLocation: ".dependencies/slackBotFunction",
       environmentVariables: {
         "RAG_MODEL_ID": props.ragModelId,
-        "QUERY_REFORMULATION_MODEL_ID": props.queryReformulationModelId,
+        "REFORMULATION_MODEL_ID": props.reformulationModelId,
         "KNOWLEDGEBASE_ID": props.knowledgeBaseId,
         "LAMBDA_MEMORY_SIZE": LAMBDA_MEMORY_SIZE,
         "SLACK_BOT_TOKEN_PARAMETER": props.slackBotTokenParameter.parameterName,
         "SLACK_SIGNING_SECRET_PARAMETER": props.slackSigningSecretParameter.parameterName,
         "GUARD_RAIL_ID": props.guardrailId,
         "GUARD_RAIL_VERSION": props.guardrailVersion,
         "SLACK_BOT_STATE_TABLE": props.slackBotStateTable.tableName,
-        "QUERY_REFORMULATION_PROMPT_NAME": props.reformulationPromptName,
+        "REFORMULATION_RESPONSE_PROMPT_NAME": props.reformulationPromptName,
         "RAG_RESPONSE_PROMPT_NAME": props.ragResponsePromptName,
-        "QUERY_REFORMULATION_PROMPT_VERSION": props.reformulationPromptVersion,
+        "REFORMULATION_RESPONSE_PROMPT_VERSION": props.reformulationPromptVersion,
         "RAG_RESPONSE_PROMPT_VERSION": props.ragResponsePromptVersion
       }
     })

packages/cdk/resources/RuntimePolicies.ts

@@ -13,7 +13,7 @@ export interface RuntimePoliciesProps {
   readonly dataSourceArn: string
   readonly promptName: string
   readonly ragModelId: string
-  readonly queryReformulationModelId: string
+  readonly reformulationModelId: string
   readonly docsBucketArn: string
   readonly docsBucketKmsKeyArn: string
 }
@@ -32,7 +32,7 @@ export class RuntimePolicies extends Construct {
       actions: ["bedrock:InvokeModel"],
       resources: [
         `arn:aws:bedrock:${props.region}::foundation-model/${props.ragModelId}`,
-        `arn:aws:bedrock:${props.region}::foundation-model/${props.queryReformulationModelId}`
+        `arn:aws:bedrock:${props.region}::foundation-model/${props.reformulationModelId}`
       ]
     })
 

packages/cdk/resources/VectorKnowledgeBaseResources.ts

@@ -156,15 +156,12 @@ export class VectorKnowledgeBaseResources extends Construct {
     // Create S3 data source for knowledge base documents
     // prefix pointed to processed/ to only ingest converted markdown documents
 
-    const chunkingConfiguration = {
-      ...ChunkingStrategy.HIERARCHICAL_TITAN.configuration,
-      hierarchicalChunkingConfiguration: {
-        overlapTokens: 60,
-        levelConfigurations: [
-          {maxTokens: 1000}, // Parent chunk configuration,
-          {maxTokens: 300} // Child chunk configuration
-        ]
-      }
+    const chunkingConfiguration: CfnDataSource.ChunkingConfigurationProperty = {
+      ...ChunkingStrategy.SEMANTIC.configuration,
+      fixedSizeChunkingConfiguration: {
+        maxTokens: 512,
+        overlapPercentage: 25
+      } satisfies CfnDataSource.FixedSizeChunkingConfigurationProperty
     }
 
     const hash = crypto.createHash("md5")
@@ -183,9 +180,6 @@ export class VectorKnowledgeBaseResources extends Construct {
           bucketArn: props.docsBucket.bucketArn,
           inclusionPrefixes: ["processed/"]
         }
-      },
-      vectorIngestionConfiguration: {
-        chunkingConfiguration: chunkingConfiguration
       }
     })