AI Theatre Chatbot — RAG Pipeline with AWS Bedrock & FAISS

Ambassador Theatre Group runs 80+ theatres across the UK. Their customer support team was drowning in the same questions over and over. What shows are on in Manchester this weekend. Is there anything good for kids in London. Whats playing at the Apollo Victoria in March. These are real questions with real answers sitting in their show catalogue, but somebody had to manually look them up and respond every single time. I built a chatbot that handles roughly 80% of these inquiries automatically.

The core is a RAG pipeline. Step Functions trigger hourly, pulling the full show catalogue from the Bolt API, enriching it with show details, chunking the data into 1,500-character documents, generating embeddings via Bedrock, and building a FAISS vector store that gets uploaded to S3. When a user asks a question, the system loads the FAISS index, runs a top-150 similarity search, then filters results by location and date. If someone asks about shows in Edinburgh next weekend, they get Edinburgh shows for those dates, not a random dump of everything in the catalogue.

Before anything touches the LLM, AWS Comprehend strips PII from the query. Names, email addresses, phone numbers, all redacted. GDPR compliance is not optional when you are processing user messages at scale. After PII redaction, the question gets summarized to extract intent, location, and date parameters. Then a 3-way classifier routes it: is this a show info request, a recommendation request, or something else entirely. Each path gets different prompt engineering and different context injection.

Claude v3.5 on Bedrock generates the final response with fit scores from 0 to 100 for each recommendation. The chatbot remembers conversation context via DynamoDB, so follow-up questions work naturally. The whole thing is serverless on Lambda, API Gateway, DynamoDB, S3, and Step Functions, deployed with Serverless Framework and Terraform. TypeScript for the chat handler and orchestration, Python for the embeddings pipeline. 80+ venues, hourly data freshness, and roughly 80% of inquiries handled without a human touching them.