A production RAG pipeline for a UK benefits platform
I own the knowledge-base pipeline end to end: ingestion and embedding on Azure Functions, vector search in Cosmos DB, answers served by an OpenAI-backed FastAPI service.
Outcome
- <3 s to answer
- 500+ documents indexed
- Hourly incremental ingest
Stack
Ask me about this
The job
Kinly is a UK benefits platform. Its team needs answers grounded in a large, changing knowledge base, fast enough to use while talking to someone. I own the pipeline that makes that possible, from the moment a document changes to the moment an answer comes back.
What I built
- Ingestion and embedding on Azure Functions, writing into Azure Cosmos DB, which also serves the vector search.
- An OpenAI-backed FastAPI service that retrieves the right passages and answers in under 3 seconds over 500+ documents.
- Scheduled ingestion for freshness: an hourly incremental ingest, plus a weekly full resync and external web crawl, staggered into the lowest-traffic window.
Pick a question
500+ documents
hourly incremental ingest
Azure Functions
ingest + embed
Cosmos DB
vector search
FastAPI + OpenAI
grounded answer
Answered in under 3 s
Illustration. In production: 500+ documents, answers in under 3 s, ingest hourly with a weekly full resync.
Also at Kinly: forms that fill themselves
I shipped eligibility-form prepopulation from website enquiries through HubSpot and Cosmos DB: it triggers when an agent opens the form, maps fields through configuration across 8 field groups, and uses freeze rules so automation never overwrites an edit an agent made.