The analysis of environmental policies usually requires the exploration of a massive multilingual corpus of heterogeneous formats, jurisdictions and institutional portals with no unified access point. Despite recent rapid progress in Retrieval-Augmented Generation (RAG) pipelines and LLM-assisted document processing, current methods do not completely address multilingual content while ensuring legal compliance and maintaining structured domain taxonomies over an incrementally growing database. This thesis presents a two-stage, open-source system for the automated data ingestion and natural language querying of around 700 climate policy documents in twenty-one different languages, generating a vector knowledge base of more than 360,000 semantically structured chunks. The first stage integrates a traffic-light license validation model, a combination of multilingual heuristic pattern matching, trusted domains boosting and the support of LLM-based classification via Mistral-Nemo (Ollama), with recursive directory exploration, with targeted file filtering across institutional portals, automated PDF quality gating, and LLM-assisted metadata extraction into a predefined glossary of hazard types, resilience criteria and other characterizing fields.