← WORKSHEET 01 / 072026LIVE

Sherpa

The assistant in the corner of this site. Ask it anything about my work and it answers from my actual projects, resume, and blog posts, rather than from a model's imagination. A retrieval-augmented (RAG) chatbot I built end to end.

PRIVATE · NO PUBLIC URL

Role
Solo build
Backend
FastAPI · pgvector
Frontend
Vue · Nuxt
Type
RAG · streaming
Client
Personal project
Timeline
2026

47

grounded in my own work, not guesses

CONTEXT

Sherpa is the chat button in the corner of this site. A visitor, often a recruiter, can ask it anything about me, my experience, my projects, my writing, and get a straight answer with sources, instead of scrolling to find it.

It answers only from my own content: the same projects and resume data that build this site, plus every blog post. Nothing about me is invented.

THE HARD PART

A chatbot on a personal site is a trap if it hallucinates. It cannot invent a job I never had or a skill I do not have, and it must not wander off into whatever a visitor asks. It also sits on a public endpoint that calls a paid API, so it has to be cheap and abuse-resistant.

So the real work was not "call an LLM", it was grounding, scoping, and hardening: make every answer traceable to my content, refuse everything else, and make sure a stranger cannot run up the bill.

WHAT I BUILT

  • A build step turns my projects, resume, and blog into one corpus, which is chunked, embedded, and stored in pgvector on Postgres.
  • On a question, the closest chunks are retrieved and the model answers only from them, citing the sources; the answer streams back token by token.
  • A FastAPI service behind nginx, deployed with Docker alongside my other projects; the browser talks to it same-origin so the API key never leaves the server.

REVISIONS · THE CALLS I MADE

A

Retrieval, not fine-tuning

What is true about me changes: a new project, a new post. RAG means I update the bot by editing content, never by retraining, and every answer stays grounded in and traceable to my own content, so it is checkable rather than a black box.

B

A relevance gate plus a scoped persona

Two layers keep it honest. If retrieval finds nothing relevant it declines instead of guessing, and the persona keeps it answering as a guide to my work and nothing else. A public bot that will confidently say anything is worse than no bot.

C

pgvector on Postgres, not a separate vector database

The corpus is small and I already run Postgres, so the vectors live next to everything else instead of adding a new system to operate, with room to grow if the content does.

D

Guards before the model call

It is a public endpoint spending real money per request, so per-IP rate limiting and an input-length cap run before any paid call, output tokens are capped, and a hard monthly spend cap sits behind all of it.

OUTCOME

  • Live on this site: ask it anything about my work.
  • Answers are grounded in my real content; it says it does not know rather than inventing.
  • Public-safe: per-IP rate limiting, input and output caps, and off-topic refusal.