@absolutejs/discover

Find the right person to contact at a company — the decision-maker PDL/Apollo charge per person-search for. Provider-agnostic orchestration: it consults importable dataset adapters first (free), then falls to LLM + web only when it needs more.

Licensed under BSL 1.1 (converts to Apache 2.0 on 2030-06-06). Use it to build your product; you may not host it as a competing prospecting/sales- intelligence SaaS. See LICENSE.

import { discoverContacts } from "@absolutejs/discover";

await discoverContacts(
  { company: "Acme", domain: "acme.com", roleIntent: "head of partnerships" },
  { sources: [secEdgar, gleif], search: braveSearch, extract: askClaude },
);
// → [{ fullName: "Jane Doe", title: "Head of Partnerships", confidence: 90,
//      linkedinUrl: "...", source: "...", reason: "listed as partnerships lead" }]

The dataset adapter contract

Open-data sources are importable adapters implementing one interface, so seeded public data is free and the paid LLM/web path is the fallback:

type DatasetSource = {
  name: string;
  findPeople?: (q: DatasetQuery) => Promise<NormalizedPerson[]>;
  findCompany?: (q: { name?; domain? }) => Promise<NormalizedCompany | null>;
};

@absolutejs/dataset-gleif, -sec-edgar, -github, … each implement this over a public dataset and normalize to one shape. Adapters are pure importers of public data. The self-collected, self-healing dataset (learned email patterns, the cross-member graph, outcomes) is deliberately not an adapter — it stays private; it's the product.

Bring your own LLM + web

discover stays provider-agnostic — you inject the capabilities:

type DiscoverDeps = {
  search?: (query: string) => Promise<WebSearchResult[]>; // Brave, SerpAPI, RAG…
  extract?: (prompt: string) => Promise<string>;          // any LLM completion
  sources?: DatasetSource[];                               // importable adapters
};

With no deps it returns only what the adapters provide. With search + extract it discovers from the open web (the LLM is told to use only people who appear in the results — never to invent names). Results are deduped and ranked by confidence. Never throws — a failing source / search / LLM just contributes nothing.