DeepQuest: A Deterministic Knowledge Graph Construction and Adversarial QA Generation System Without Large Language Models

DeepQuest is a fully deterministic deep web research engine that does not rely on any large language models (LLM). It constructs knowledge graphs through structured data extraction and automatically generates adversarial multi-hop question-answer pairs. This system aims to address the hallucination issues, high computational costs, and behavioral uncertainty of LLMs, providing a new path for scenarios requiring high interpretability and controllability such as legal document analysis and medical knowledge base construction.

In recent years, LLMs have dominated fields such as knowledge extraction and question-answering systems, but they face issues like hallucinations, high costs, and behavioral uncertainty. DeepQuest proposes a completely different approach: building a fully deterministic research engine that enables deep web crawling, knowledge graph construction, and adversarial QA generation without using LLMs, which is of great significance for scenarios requiring high interpretability.

DeepQuest's core architecture consists of three deterministic modules:

1. Deep Web Crawling: Based on a deterministic parsing strategy using structured tags, it identifies entities and relationships, featuring predictability, auditability, and low resource consumption;
2. Knowledge Graph Construction: Stores facts in triple form, establishes entity links through ontology mapping and pattern matching, with processes including entity recognition and disambiguation, relationship extraction and verification, and graph fusion and deduplication;
3. Adversarial Multi-hop QA Generation: Traverses multi-hop paths in the knowledge graph to generate complex question-answer pairs requiring reasoning chains (e.g., the multi-hop reasoning path from Einstein to the photoelectric effect).

DeepQuest's technical advantages include:

• Full Determinism: The same input always produces the same output, suitable for scenarios like regulatory compliance review, scientific literature organization, and multi-team collaboration;
• No LLM Dependency: Achieves or exceeds LLM accuracy in specific domains through rule engines and pattern matching, avoiding API costs, hallucination errors, and black-box issues;
• Adversarial Testing Capability: The generated multi-hop QA pairs can serve as evaluation benchmarks for systems like RAG, testing multi-hop reasoning and factual accuracy.

DeepQuest's application scenarios include:

• Enterprise Knowledge Management: Builds private knowledge graphs, supports precise retrieval, and runs sensitive data locally;
• Academic Research Assistance: Automatically organizes literature reviews, constructs domain knowledge networks, and discovers research connections and gaps;
• Model Evaluation Benchmark: Adversarial QA pairs are used to evaluate the multi-hop reasoning capabilities of LLMs and RAG systems.

DeepQuest has the following limitations:

• Domain Adaptability: The rule engine needs to be reconfigured for new domains;
• Unstructured Content: For highly unstructured text, pattern matching may not perform as well as LLMs;
• Maintenance Cost: Parsing rules need to be continuously updated when web content formats change.

DeepQuest represents a technical exploration returning to deterministic methods. While embracing the convenience of LLMs, it emphasizes the value of structure, determinism, and auditability. It provides an alternative for scenarios requiring interpretability, consistency, and cost-effectiveness. As AI moves toward production, the demand for system reliability and predictability increases, and DeepQuest's philosophy may become a core trait of the next generation of knowledge systems.