Ask Me Anything — RAG Widget

Shipped

A RAG-powered chat widget that answers recruiters’ questions about my background in natural language.

PythonFastAPIPineconeClaude APIReactAWSCloudflare

Visitors ask natural-language questions about my professional background and get grounded, conversational answers — the model answers about me, in the third person, citing its sources. Built across three fully-designed layers: Notion ingestion, a Python/FastAPI query pipeline, and a React streaming frontend.

Highlights

Ingestion: Notion → chunk → hosted Pinecone embedding (llama-text-embed-v2 @ 384-dim, cosine) → vector store.
Query pipeline: retrieve → relevance gate → grounded prompt → Claude Haiku 4.5, streamed over SSE from FastAPI.
Answers are grounded and cited — a relevance gate keeps the model from answering off-corpus questions.
Cloud rollout on AWS Lightsail + Cloudflare, with Lambda + EventBridge for scheduled ingestion.
The frontend widget is a first-party, style-isolated component embedded directly in this site at /ask.

Architecture

The widget is split into independently designed layers so each can be reasoned about and tested on its own: a Notion ingestion pipeline, a Python/FastAPI retrieval-and-serving pipeline, and a React streaming frontend. Embedding is a hosted API call by design — there is deliberately no local torch / sentence-transformers dependency.

Screenshots

A full ask-me-anything cycle — type a question, watch the answer stream in token by token, source cards appear, then an off-topic question is politely declined by the relevance gate. — One full cycle: ask a question, watch it stream in, see the sources — then an off-topic question is turned away by the relevance gate.

The idle Ask Me Anything widget with an empty question box and suggested-question chips. — The idle widget — an empty ask box with suggested questions to get started.

The widget rendering an answer as it streams in token by token from the FastAPI backend over SSE. — An answer streaming in token by token, served over SSE from FastAPI.

Source cards below a completed answer, grouped by portfolio page with a preview snippet and a read-more link. — Every answer is grounded — the source cards show the portfolio content it drew from.

An off-topic question answered with a polite decline and no source cards, showing the relevance gate at work. — Ask something off-corpus and the relevance gate declines — no sources, no hallucinated answer.

Repository

njranum/ama-rag