gem-hunter
主要言語: TypeScript
Two-agent Greenhouse job search: one writes Google X-ray queries or rejects, you curate them, the other summarises what came back. Next.js + Vercel AI SDK + zod.
Want to build this with me? I'm running a 2-day workshop on building agents and setting up a RAG pipeline. Sign up here.
Type what you want. One agent writes boolean queries (or rejects the request), you pick which to run, they run, and a second agent reads the postings and tells you which are worth opening.
"ai engineer, llm and rag"
│
▼ searchAgent reject, or 1–5 boolean queries
│
▼ you remove or edit any, then run
│
▼ Firecrawl search Greenhouse, pages come back as markdown
│
▼ searchSummaryAgent picks, why, and what the results missed
Three accounts. Two are free; OpenAI needs a few dollars of credit. Ten minutes total.
The agents run on gpt-5.4-mini and gpt-5.4.
OPENAI_API_KEY=sk-...
In the workshop you get a key for the Parsity proxy instead, plus
OPENAI_BASE_URL=https://parsity-litellm.fly.dev/v1. Same code, no OpenAI account needed. Use unprefixed model names on the proxy —openai/...aliases return "invalid model ID".
One call searches Google scoped to Greenhouse and returns every result page already cleaned to markdown.
FIRECRAWL_API_KEY=fc-...
Free tier: 500 credits and 10 requests a minute. A query with scraping costs about 10 credits, so that is ~50 queries a month — and each query is one request against the per-minute limit. Use your own key; you cannot share one with a room.
Every model call traced: the prompt it got, the tokens, the cost, the latency. Optional, but you will not understand why the agent did something without it.
LANGSMITH_TRACING=true
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_API_KEY=lsv2_...
LANGSMITH_PROJECT=greenhouse-xray
LANGSMITH_ENDPOINT is required — without it nothing is sent, silently.
cp .env.example .env.local # paste the keys in npm install && npm run dev # http://localhost:3000
Everything is in lib/agents.ts. An agent is a function that calls a model
with a schema and some examples:
export const searchAgent = (messages) => generateText({ model: model('gpt-5.4-mini'), output: Output.object({ schema: planSchema }), system: `Respond as in these examples.\n\n${shotsBlock}`, messages, });
Schema first. Never ask for text you then have to parse. Ask for the shape you want back and let zod refuse anything else:
const planSchema = z.object({ action: z.enum(['search', 'reject']), reason: z.string(), // making it explain improves the answer queries: z.array(z.string()).min(1).max(5), });
Examples, not prose. The search agent has no system prompt. PLAN_SHOTS is
a list of { request, output } pairs and each reason states the rule that
example exists to teach. When the agent misbehaves, add an example — do not
add a paragraph. Watch one example over-generalise: the "we do not hire
juniors" shot also rejects "entry level".
Reject before you spend. The first agent runs on the small model and decides whether to do anything at all. Junk, injection, and "who is the CEO" stop there for a fraction of a cent.
Human in the loop. Queries go to the screen before they run. That is the cheapest correction in the whole system.
Trust nothing back. The summary agent returns URLs; anything not in the pages it was actually given is dropped before render.
npm run eval with the dev server running. A case is a request, the action it
should take, and a rubric a second model grades against:
{ request: 'senior backend engineer, golang', action: 'search', rubric: 'Titles cover backend engineering; keywords carry golang and Go.' }
The judge is not deterministic — read the reasoning, not just the score.
main runs, with four TODO(n) markers to fill in:
fixed has 1, 3 and 4 done. Follow-up questions are the workshop exercise.