GeoMind

Already contacted

Addressing the academic hallucination (fabricating papers and citations) and unreliable information sources (unverifiable) problems common in existing AI research tools (Gemini Deep Research, Claude Research), GeoMind proposes a multi-agent literature review solution based on a trustworthy knowledge base of 145K real papers (CAS Zone 1 journals, DOI traceable), eliminating hallucinations at the source. Collaboration scenario: Planner decomposes topics and dispatches queries → Research retrieves from trustworthy library → Critic reviews relevance, rejects if substandard triggering supplementary search → Writer generates review after quality passes. Four Agents form 'retrieval-review-iteration-output' loop ensuring every citation is verifiable.

System contains 4 Agents: Planner (coordinator) receives user topics and decomposes task scheduling; Research (retriever) performs semantic retrieval from 145K paper library; Critic (reviewer) evaluates literature relevance and quality gates; Writer generates outlines and complete reviews. OpenAgents collaboration: All Agents inherit WorkerAgent base class for unified management; Channel broadcast (#geomind-review) publishes progress notifications for real-time user viewing; Direct Message point-to-point communication for structured JSON task dispatch and result submission; Event system (on_channel/on_direct) responds to messages for async collaboration; Workspace unified management of Agent message routing.

① Planner Agent: Receives user research topics (supports Chinese/English), auto-translates and generates multi-angle search keywords, coordinates overall process. ② Research Agent: Semantic retrieval from 145K CAS Zone 1 paper library, based on Qdrant vector search + FastEmbed local embedding, returns DOI-traceable real literature. ③ Critic Agent: Reviews retrieval results' relevance to topic, calculates quality score; rejects and triggers supplementary search if score <7, forming quality loop. ④ Writer Agent: Generates outline based on approved literature set, calls DeepSeek LLM to write complete review, ensuring every citation is verifiable. ⑤ Agent collaboration: Through OpenAgents Direct Message for structured task passing (JSON), Channel broadcasts progress status, Event system (on_direct/on_channel) enables async response. ⑥ Implements Qdrant Search Mod: Encapsulates vector retrieval capability, supports year filtering and relevance reranking.