Architecture Overview · Park Whisperer Platform

Natural Language Operations Assistant

A conversational AI agent that auto-routes questions to the right knowledge domain, conditionally enables live tools, and generates grounded conversational replies via Claude Sonnet with Bedrock Prompt Cache — API key never exposed to the browser.

● Production · Auto-RoutingClaude Sonnet · Bedrock CacheSWA Proxy Auth

Click any component to explore it in the full case study.

Input A
User Natural Language Query
Park operations question or visitor inquiry. Supports implicit pipeline routing — keywords in question automatically determine content_type and knowledge_types without explicit user selection.
Natural Language · Auto-routed
Input B
Conversation History
Last 4 turns (500 chars each) injected for multi-turn context. SWA proxy routes /api/pipeline/proxy/api/agentChat → Function App with API key injected server-side.
4-turn history · SWA proxy auth
Query Processing + Retrieval + Generation
Step 1
Pipeline Routing
_pipeline_from_question() keyword-matches → content_type assignment → knowledge_types filter for retrieval
›
Step 2
Live Tool Check
_needs_live_tools() detects historical vs. live keywords. Conditionally enables tool set to avoid live-data spam on historical queries.
›
Step 3
SLM Classification
Phi-3 Mini / Phi-4-mini determines retrieval strategy (semantic vs. structured vs. point-read)
›
Step 4
RAG Retrieval
PostgreSQL HNSW semantic search + Cosmos DB structured queries. Context assembled.
›
Step 5
Claude Generation
Claude Sonnet 4.5 + Bedrock Prompt Cache (cache when context > 1200 chars). Conversational tone enforced.
🔒
Anti-Hallucination + SWA Proxy Auth
Same data-source-only rule as pipelines — never fabricate. Historical keyword detection prevents unnecessary live tool calls. API key never exposed to client (SWA proxy injects). Bedrock Prompt Cache reduces cost on multi-turn.
Data-source-only ruleHistorical keyword filterSWA proxy key injectionBedrock Prompt Cache (>1200 chars)No API key on clientTool call trace in UI
Output A
Conversational Reply
1–3 paragraph plain text response. No section markers, no markdown headers, no fabricated data. Tool call trace available for debug display in UI.
Plain text · No markdown
Output B
Token Usage Metadata
Cost tracking per query. Input/output token counts, cache hit/miss status, model ID used. Supports optional staging/prod backend override via URL param or localStorage.
Cost tracking · Backend override
Auto
Pipeline routing
15+
Live tools available
4
History turns
Cache
Bedrock Prompt Cache
StackPythonClaude Sonnet 4.5AWS BedrockBedrock Prompt CacheLangChainPostgreSQL HNSWCosmos DBPhi-3 Mini SLMAzure SWA proxy