We built an AI system for a bank that answers natural-language questions about its corporate and regulatory knowledge base. It combines external regulations with internal regulations, policies, instructions, and templates, and returns a structured answer with requirements, related documents, and citations to primary sources.
Classic semantic search is not enough here. An answer has to account for the links between regulations, internal documents, and requirements, pull data from several sources, and stay verifiable: every conclusion rests on a specific primary source.
The solution combines RAG with a corporate knowledge graph. The system breaks uploaded PDF, Word, and ZIP files into sections, articles, clauses, and requirements, and links the extracted entities to each other. For each query, the graph is rebuilt around the relevant topic, and the answer draws on three sources at once: vector search, knowledge graph links, and the LLM. Five scenarios run on top of this: finding applicable regulations, requirements mapping, document review, comparing internal regulations with external requirements, and checklist-based legal screening.
- 0101
Knowledge base analysis
Studying the structure of laws and regulations, internal regulations, policies, instructions, and templates.
- 0202
Document pipeline
Building ingestion and parsing for PDF, Word, and ZIP files and extracting text and document structure.
- 0303
Knowledge graph
Extracting entities and linking documents, requirements, regulations, processes, and sources.
- 0404
RAG retrieval
Implementing vector search, the retrieval pipeline, and natural-language answers with source citations.
- 0505
Document review
Scenarios for finding missing requirements, potential conflicts, and weak wording.
- 0606
Interface & testing
Knowledge base interface, link visualization, and testing the system on real use cases.
- 01LLM
- 02RAG
- 03Knowledge Graph
- 04Vector Search
- 05OCR
- 06Python
- 07FastAPI
- 08PostgreSQL
- 09Graph DB
- 10LangChain / LlamaIndex
Have a similar challenge?
We’ll review it and estimate within 5 business days.