01. Cognitive Decoupling
Standard Large Language Models (LLMs) often conflate retrieval and reasoning, leading to the well-documented phenomenon of "hallucination." GCSE Tutor solves this through **Cognitive Decoupling**.
When a query is received, it is first processed by the **Investigation Loop**. This agent-cluster is prohibited from generating answers; its sole mandate is to scour the Knowledge Graph, local PDFs, and official exam board specifications via MinerU parsing to establish a "Grounded Truth."
Investigative Phase
RAG Hybrid retrieval scours vector embeddings and entity-relation mappings to find the exact mark-scheme criteria.
Reasoning Phase
The Solve Loop uses Chain-of-Thought (CoT) to build a multi-step mathematical or logical solution based ONLY on the investigative data.
02. Grounded Retrieval (RAG)
Our **Grounded Retrieval** system utilizes a hybrid approach. While vector search identifies semantic similarity, our **Knowledge Graph (LightRAG)** integration maps the intricate relationships between concepts—such as the link between "Mitosis" and "Cell Differentiation" in AQA Biology Paper 1.
The MinerU Pipeline
Official PDF specifications are parsed via MinerU into high-fidelity markdown. This ensures that every citation generated by the system refers to a verified, board-compliant source.
03. Verification Protocol
The final stage of the architecture is the **CheckAgent**. This specialist agent is tuned specifically to mark scheme rigidity. It acts as a "Digital Examiner," reviewing the output from the Solve Loop and cross-referencing it against board-specific keywords.
The A* Guard
"If the solution does not meet the 99% precision threshold for the target grade, the system automatically triggers a recursive re-investigation. We do not settle for 'good enough'."