1. Architecture Diagram

flowchart TD A[Sherlock Holmes novels TXT / EPUB / PDF] --> B[S3 Raw Data Bucket] B --> C[Ingestion Lambda / ECS Task] C --> D[Text Cleaning + Chunking] D --> E[Embedding Model Amazon Bedrock Titan / Cohere] E --> F[Vector Store OpenSearch / Aurora pgvector] U[User] --> G[Web App React / Next.js] G --> H[API Gateway] H --> I[Chat Lambda / FastAPI on ECS] I --> J[Retrieve Relevant Chunks] J --> F J --> K[Prompt Builder] K --> L[LLM Claude / Amazon Nova / Llama via Bedrock] L --> M[Answer with citations] M --> G I --> N[CloudWatch Logs] B --> O[Optional: Metadata DB DynamoDB]

2. Problem Statement

Build an AI chatbot that allows users to “talk to” the Sherlock Holmes canon.

Instead of manually searching long novels, users can ask natural-language questions such as:

“What does Holmes think about observation?”

“Summarise Irene Adler’s role.”

“Which cases involve poison?”

“Answer in Holmes’ tone, but cite the story.”

The system solves:

Knowledge access: Converts long literary text into searchable semantic knowledge.

Conversational UX: Lets users explore literature through dialogue.

Grounded generation: Uses RAG so answers are based on retrieved book passages, reducing hallucination.

Portfolio value: Demonstrates LLM integration, vector search, AWS architecture, cost trade-offs, and AI product thinking.

3. Architecture Decisions

To Research

4. Trade-offs

Cost vs Performance

To Research

Latency vs Complexity

To Research

Latest MVP Scope

Ingest public-domain Sherlock Holmes text into S3.
Clean and chunk text by story/chapter.
Generate embeddings using Bedrock.
Store vectors in PostgreSQL pgvector.
Build a Lambda API:
Build a simple web chatbot.
Add evaluation questions to test answer quality.

Carina's Blog

1. Architecture Diagram

2. Problem Statement

3. Architecture Decisions

4. Trade-offs

Cost vs Performance

Latency vs Complexity

Latest MVP Scope