1 · Retrieve
Two kinds of search
BM25 and vector similarity, fused with reciprocal rank fusion, over a validated index.
Projects / Production RAG Assistant
The assistant on this site answers questions about my work from published evidence. It runs on AWS and Cloudflare, with rate limits and a budget reserved before each paid call.
You work with the person who built it. The first 30 minute call is free, and if I can’t help, I say so before you pay anything.
StatusLive
Answer quality depends on production behavior too: latency, citations, rate limits, monitoring and cost control.
“Will it make things up?”
It retrieves published evidence first, generates once, then verifies protected claims before committing the final answer.
“Could someone run up the bill?”
The worst case is reserved against daily and monthly caps before any paid call. If the budget store is down, it fails closed.
“Could it be abused?”
A Cloudflare Worker and firewall, IP rate limits, body size limits and a kill switch, plus fixed rate windows in the app.
“Will people sit waiting for an answer?”
Drafts stream as they’re written and stay provisional. Only the verified final answer is committed.
Answer quality depends on production behavior too.
Published evidence first, one generation, and a check before anything is final.
1 · Retrieve
BM25 and vector similarity, fused with reciprocal rank fusion, over a validated index.
2 · Generate
System prompt, evidence and the visitor’s question produce a draft, shown as provisional.
3 · Verify
Protected claims are verified before the final answer replaces the draft.
The path from a visitor’s question to a verified answer.
browserCloudflare Pages chat UIedgeCloudflare Worker: WAF, rate limits, kill switchfunction URLAWS Lambda with response streamingappFastAPI: rate limit, budget reservationretrievalBM25 + vector, fused with RRFindexversioned S3 bundle, loaded on cold startstreamsearching, token, final or error
AWS CDK for infrastructure, and GitHub Actions for offline tests, security checks and deploys with a live smoke test. Failure metrics and alarms go to CloudWatch.
Visitor prompt text isn’t written into metric dimensions. Accepted questions and answers may be kept for about 30 days for debugging.
The first 30 minute call is free, and I reply within one business day.