Enterprise Knowledge Base Assistant Using Multi-Model GenAI Architecture

accenture · Healthcare · 2023 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Public-health knowledge assistant

Accenture and AWS built Knowledge Assist to help a large public health client provide up-to-date, multilingual, self-service answers over unstructured knowledge content while reducing call-center escalation and improving citizen access.

Constraints:

• The client serves millions of citizens every day.
• Knowledge changes frequently in a health landscape.
• Traditional search and FAQ bots were too cumbersome and narrow.
• The system needed multilingual support, citations, conversation memory, and integration with an existing web platform.
• Model selection had to consider supported use cases, maximum tokens, cost, accuracy, performance, and languages.

Architecture

Knowledge Assist combines Amazon Lex, Lambda orchestration, DynamoDB conversation storage, Amazon Kendra retrieval, Amazon Bedrock with Claude-2 for generation/translation, Amazon Titan embeddings, Pinecone vector search, S3/Route 53/CloudFront front-end hosting, CloudWatch logs, OpenSearch, and Kibana reporting.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  cloudfront-s3-front-end["CloudFront S3 front end"] --> amazon-lex["Amazon Lex"]
  amazon-lex --> lambda-orchestrator["Lambda orchestrator"]
  lambda-orchestrator --> dynamodb-conversations["DynamoDB conversations"]
  lambda-orchestrator --> amazon-kendra["Amazon Kendra"]
  lambda-orchestrator --> bedrock-claude-2["Bedrock Claude-2"]
  bedrock-claude-2 --> amazon-titan-embeddings["Amazon Titan embeddings"]
  amazon-titan-embeddings --> pinecone-vector-database["Pinecone vector database"]
  pinecone-vector-database --> lambda-orchestrator
  lambda-orchestrator --> cloudwatch-opensearch-reporting["CloudWatch OpenSearch reporting"]

Components

Data flow

A user submits a request through the CloudFront/S3 front end; Amazon Lex detects intent and calls a Lambda orchestrator; Lambda reads conversation state from DynamoDB, gets Kendra/Pinecone retrieval context, sends the prompt to Bedrock Claude-2, post-processes the response, and logs interactions to CloudWatch/OpenSearch.

Why this architecture

Accenture combined conversational intent routing, RAG retrieval, Bedrock generation, vector embeddings, and reporting because the client needed broader natural-language access, citations, multi-turn support, and multilingual answers for a public-health knowledge base.

Models used: Amazon Bedrock · Claude-2 · Amazon Titan · Amazon Kendra · Amazon Lex · Pinecone · DynamoDB · Lambda

Scale

• The public-health client serves millions of citizens every day.
  “A large public health sector client serves millions of citizens every day”article, stated
• Knowledge Assist is reported to reduce training time for new hires by over 50%.
  “reduces training time for new hires by over 50%”article, stated
• The solution is reported to cut escalations by up to 40%.
  “cuts escalations by up to 40%.”article, stated
• Kendra retrieves the top five relevant search results for prompt context.
  “get the top five relevant search results to build the relevant context.”article, stated

Alternatives considered (roads not taken)

Use the traditional FAQ/chatbot framework unchanged.

Rejected because: The existing experience was narrow and cumbersome compared with natural conversational access.

Choose a foundation model without regression testing.

Rejected because: Accenture tested Bedrock models for use-case fit, tokens, cost, accuracy, performance, and languages before choosing Claude-2.

Return answers without citations.

Rejected because: The listed features include transparent citations guiding users to source information.

Patterns

rag-retrieval (retrieval)

Knowledge Assist retrieves Kendra and embedding-context results before sending prompts to Claude-2.

model-routing (model-ops)

Accenture regression-tested multiple Bedrock model providers before selecting Claude-2.

observability-tracing (operations)

The architecture logs request/response metadata through CloudWatch into OpenSearch for reporting.

Tradeoffs

Integrate GenAI into an existing FAQ bot.

Chose Augment the existing chatbot with retrieval and Bedrock generation. over Replacing the whole web interaction model..

Rationale: The client needed broader questions answered while preserving self-service access.

Use Claude-2 via Bedrock.

Chose Claude-2 selected after model regression testing. over Unvalidated model/provider choice..

Rationale: Claude-2 best matched the tested use-case, token, cost, accuracy, performance, and language requirements.

Gotchas (what bit them)

Traditional knowledge bases remain cumbersome despite content coverage.

Root cause: Users still struggle to locate the exact information they need through search or FAQ interfaces.

Resolution: Use a conversational RAG interface with citations and context-aware answers.

Model choice must account for production constraints.

Root cause: Supported use cases, token limits, cost, accuracy, performance, and languages all matter.

Resolution: Regression-test candidate Bedrock models and select the best fit.

Multi-round conversations require explicit state.

Root cause: The chatbot needs session and user history to maintain context.

Resolution: Save conversations in DynamoDB application databases.

Evaluation

Accenture evaluated models by regression tests across model attributes and reports business outcomes including training-time reduction and escalation reduction.

Metrics: Supported use cases · Maximum tokens · Cost · Accuracy · Performance · Languages · Training time over 50% reduction · Escalations up to 40% reduction

Outcomes

• Knowledge Assist reportedly reduces new-hire training time by over 50%.
• The solution reportedly cuts escalations by up to 40%.
• The platform supports multilingual, citation-backed, conversational answers over enterprise content.

The public-health client serves millions of citizens daily, and the architecture is presented as reusable across health sciences, financial services, manufacturing, and more.