Liepin Agentic Sourcing Pipeline

Problem

Candidate sourcing, resume screening, and outreach on Liepin are often fragmented across browser sessions and manual judgment. Recruiters may open dozens of profiles, compare signals inconsistently, and lose track of the criteria they started with.

How It Works

This skill packages Liepin sourcing into a background pipeline. First, create a criteria.json with explicit fields such as position_name, target_count, position_scope, cities, seniority, and must_have_signals; city and seniority must be provided rather than inferred. After submission, the JD-Worker generates a keyword search plan and can operate in initial, widen, or tighten modes. The Browser-Worker uses a saved Liepin session to run searches, collect candidate previews, harvest detail pages, and click “立即沟通” when appropriate. The Resume-Worker evaluates candidates with an LLM and promotes promising profiles into a pending_communication state. The Orchestrator adapts the search: tighten immediately when results exceed 200, drain the current pool when target count is unmet, and cap iterations at five. Preview collection can paginate through up to five pages, about 100 candidates, while detail harvesting and communication share the same page session to avoid repeated login or navigation.

Limits And Notes

The skill requires Liepin credentials and two LLM configurations in .env, plus a saved valid Liepin session before first use. It also handles known page quirks such as the full-screen “AI 帮搜” overlay and compound keyword inputs. It is intended for engineers who want to script Liepin sourcing and initial screening, not for unsupervised mass outreach.