Back to selected work
Engineering, Personal project
RAG analytics assistant
An assistant that answers analytics questions from grounded retrieval, with routing and caching to control latency and cost.
- Role
- Designed and built
- Stack
- Python, Embeddings, Vector search, Caching, Evaluation
- Question
- Normalise and route
- Filtered retrieval
- Grounded answer
- Cache and evaluate
- Answers
- Grounded in retrieved data
- Releases
- Regression-checked
Problem
People wanted quick answers about their data. A plain chatbot answers confidently even when it has no grounding, and every call costs time and tokens.
Approach
- Retrieval filtering and better chunking so the model sees the right context.
- Query normalisation plus caching for repeated questions.
- Routing to skip unnecessary model or tool calls.
- A standard evaluation workflow so prompt or model changes can't silently reduce quality.
Outcome
Faster responses on repeated queries, steadier answer relevance, and controlled per-query cost.
What I'd improve
- A golden-question set run on every change.
- Guardrails and audit logs for sensitive data.