KNOWL: Retrieval-augmented generation and hybrid knowledge-graph retrieval for more reliable enterprise LLM applications — Implementation and results

Retrieval-augmented generation has become a practical response to the factual and governance limitations of purely parametric large language models in enterprise settings, but conventional document-centric retrieval remains brittle when the task requires the coordinated reuse of heterogeneous prior knowledge, multi-step information gathering, and structured document production. A preceding paper on the KNOWL project established the project context, reviewed the state of the art in retrieval-augmented generation, GraphRAG, multilingual grounding, PDF-centric processing, and governance, and formulated the hypotheses to be tested in a subsequent implementation-focused study. The present paper continues that line of work by reporting the actual technical execution of the project. It analyses the transition from the original project objectives to the implemented system, describes the architecture and workflow adopted, documents the construction of the knowledge graph and the GraphRAG pipeline, and reports the outcomes of prototype integration and pilot validation in a real business process for bid and proposal preparation in the BPO sector. The implemented solution combined a Neo4j knowledge graph, vector retrieval, open-weight language models deployed in a controlled environment, and a web application that supported project creation, RFP ingestion, requirement analysis, area-based work allocation, proposal drafting, executive-summary generation, contextual querying, and export of editable outputs. The resulting system reached TRL6 and was validated in a preproductive pilot over real RFP documentation. Reported pilot results indicate strong operational acceptance and perceived utility, including a content quality score of 9.67/10, time to draft between 20 and 30 minutes, and estimated time savings of 70% per offer.