Ask Nare

Problem

A portfolio is static; recruiters have specific questions. I wanted a widget that answers them only from what I have actually written down, refuses politely when it should, and proves it with a scorecard instead of a promise.

What I built

• Knowledge base: content/profile/*.md chunked at build time. Embeddings (Voyage or OpenAI) when a key is configured, BM25 keyword search otherwise — so it runs with only ANTHROPIC_API_KEY.
• Tools exposed to the model: search_profile(query), get_project(slug), get_availability().
• System-prompt guardrails: answer only from retrieved content, cite the section, say "I don't know" when the content doesn't cover it, decline off-topic requests, never state salary, never speculate about degree completion.
• Rate limit by IP (token bucket), 20 turns per session, capped output tokens. Latency and approximate cost under each answer.
• npm run eval:ask runs content/evals/ask-nare.jsonl through deterministic checks and an LLM judge and fails under a threshold.

Architecture

Result

See the eval scorecard. Try it on the home page.

Stack

Next.js route handler · @anthropic-ai/sdk · TypeScript · Vitest