RAG pipelines για ESG compliance: τεχνική ανάλυση
Πώς φτιάξαμε υβριδικό σύστημα ανάκτησης FAISS + BM25 που συμπληρώνει αυτόματα ερωτηματολόγια GRI από έγγραφα εταιρειών.
ESG compliance is drowning in paperwork. GRI questionnaires can have 200+ questions, each requiring evidence from company documents. We built a RAG pipeline that does it in minutes.
The Architecture
We use a hybrid retrieval approach: FAISS for semantic search and BM25 for keyword matching, merged with Reciprocal Rank Fusion. Semantic understanding for paraphrased content, exact matching for specific metrics.
Document Ingestion
Companies upload PDFs, Word docs, and spreadsheets. We extract text, chunk it into ~500 token segments with overlap, generate embeddings and store them in FAISS. The BM25 index is built from the same chunks.
Multi-LLM Support
Some clients want GPT-4o, others prefer Claude for long documents, and a few need on-premise models. An abstraction layer lets us swap the model and keep the same prompts and output format.
Impact
The platform reduced questionnaire completion time by 80%. Consultants now review and refine AI-generated answers instead of writing them from scratch.