MCP PDF Tools
800 regulatory PDFs a month. Jurisdiction matching in seconds, not half a day.
The Problem
A fintech company’s regulatory compliance team receives approximately 800 PDF documents per month: regulatory guidance, updated frameworks, national transpositions of EU directives, EBA/ESMA technical standards. Their job is to identify which documents actually affect the company’s products in which jurisdictions.
Before any analysis happens, someone has to read enough of each document to know whether it’s relevant. With 800 documents and a team of four analysts, that’s 200 documents per person per month. At 15 minutes per initial scan, that’s 50 analyst-hours per week just on triage.
What We Built
An MCP server with four primary tools targeting the triage workflow:
extract-pdf-content— full text extraction with page-level segmentation and metadatafind-related-pdfs— given a target PDF, finds semantically and textually related documents in a directorysearch-pdf-directory— regex/pattern search across hundreds of PDFs simultaneouslymerge-pdfs-ordered— assembles multiple PDFs in a specified order with optional cover page
The find-related-pdfs tool is the triage accelerator. It extracts key terms from the target document, then ranks the directory by relevance — analysts see the 12 most-related documents out of 800 in seconds.
How It Works
The relevance matching combines text pattern extraction with substring overlap scoring:
def find_related_pdfs( target_pdf: str, search_directory: str, max_results: int = 10 ) -> list[RelatedDocument]: # Extract key regulatory terms from target target_text = extract_text(target_pdf) target_terms = extract_regulatory_terms(target_text) results = [] for pdf_path in iter_pdfs(search_directory): try: candidate_text = extract_text(pdf_path, max_pages=5) # header pages score = compute_relevance(target_terms, candidate_text, pdf_path) results.append(RelatedDocument(path=pdf_path, score=score)) except Exception: continue # malformed PDFs are common in regulatory archives return sorted(results, key=lambda r: r.score, reverse=True)[:max_results] def compute_relevance( terms: list[str], candidate_text: str, candidate_path: str ) -> float: # Text overlap score term_hits = sum(1 for t in terms if t.lower() in candidate_text.lower()) text_score = term_hits / max(len(terms), 1) # Filename similarity (regulators often number related documents) filename_score = fuzzy_filename_match( Path(target_path).stem, Path(candidate_path).stem ) return 0.7 * text_score + 0.3 * filename_score The filename fuzzy matching handles a common regulatory pattern: the EBA often publishes EBA-GL-2024-01.pdf, EBA-GL-2024-01-Annex-I.pdf, EBA-GL-2024-01-Corrigendum.pdf as related documents. The fuzzy match catches these without requiring the analyst to know the naming convention.
The Outcome
The compliance team now runs find-related-pdfs against each new regulatory publication and receives a ranked list of existing documents in their archive that it relates to. This gives them instant regulatory lineage — “this EBA opinion updates these three previous guidelines” — without manual cross-referencing.
The search-pdf-directory tool was unexpectedly useful for answering internal queries: “which of our regulatory filings mention DORA Article 28?” returns results across the full archive in under 30 seconds.