Conversation needs product truth.
The assistant must understand open-ended intent while keeping recommendations anchored to approved attributes, availability and commercial rules. Unsupported answers fall back instead of becoming product advice.
03 / Independent AI reference architecture
A technical blueprint for a boutique AI sales assistant.
A personal, client-neutral reference architecture for combining conversational guidance with trustworthy product context. It is not presented as commissioned by, deployed for or representative of any employer or client.
LangGraphcontrolled agent workflow
Hybrid RAGsemantic + structured retrieval
RAGASregression evaluation
FastAPItyped service boundary
Orchestrate reasoning, retrieval and tools.
FastAPI defines the service boundary while LangGraph coordinates intent classification, retrieval, bounded tool execution and response synthesis. Bedrock-hosted models use semantic search together with structured catalogue queries.
Evaluate the system, not only the answer.
A golden test set and RAGAS-style measures cover faithfulness, relevance and retrieval quality. Citations, traces, guardrails and low-confidence fallbacks keep behaviour reviewable and human-led.
Text architecture
How a grounded answer moves through the system.
- 01User intent
- 02Policy check
- 03Hybrid retrieval
- 04LangGraph workflow
- 05Bedrock model
- 06Grounded validation
- 07Cited response or safe fallback
Technical answers
The architecture, in plain language.
What does the assistant do?
It turns open-ended product questions into grounded guidance by combining conversational intent with approved catalogue attributes, availability and commercial rules. When evidence is insufficient, the assistant explains the limitation or falls back safely.
How does LangGraph control the workflow?
LangGraph models the interaction as explicit states and transitions for intent classification, retrieval, bounded tool use, synthesis and review. That structure makes execution traceable and prevents an agent from skipping required checks.
Why combine semantic and structured retrieval?
Semantic retrieval handles meaning and discovery; structured queries preserve precision for attributes, availability and rules. Hybrid retrieval keeps the experience conversational without treating generated language as product truth.
How is the assistant evaluated and guarded?
A golden test set and RAGAS-style measures track faithfulness, relevance and retrieval quality. Citations, tool boundaries, traces and low-confidence fallbacks keep responses reviewable and human-led.
Primary technical references
My scope
Reference architecture · LangGraph orchestration · retrieval design · API contracts · evaluation · guardrails
Continue