BEBahae Eddine
back to projects
Enterprise AI · Knowledge CopilotConcept

OCP AI Assistant — Enterprise Knowledge Copilot

Concept case study with OCP Group (Energy & Oil) — an enterprise knowledge copilot that grounds every answer in the company's own documentation instead of generic model knowledge.

stack
PythonFastAPILangChainAzure AI FoundryMilvusNext.js

Concept case study developed under XAI for OCP Group — the global leader in phosphate and fertilizers, operating across energy and oil-adjacent industrial processes. The brief: employees should be able to ask questions in natural language and get answers grounded in the company's own operational and technical documentation.

Problem

Industrial groups run on knowledge that lives in engineering specs, safety procedures, SOPs, project archives, and decades of tribal expertise. Finding the right document is slow; asking a colleague is faster but unreliable; and a generic chatbot is worse than both because it invents answers. OCP needed a copilot — something that knows what the organization knows.

Approach

An enterprise knowledge copilot on a retrieval-augmented generation (RAG) architecture:

  • Enterprise corpus: technical documentation, procedures, and reports normalized into a searchable knowledge base.
  • Hybrid retrieval: dense + keyword search over Milvus, with metadata filters (site, system, document type, language) so retrieval stays relevant to the questioner's context.
  • Grounded generation: the assistant answers exclusively from retrieved passages and shows its sources — a "show me where that comes from" button is not a luxury, it's the whole point.
  • Sovereign deployment: the pipeline runs on constrained infrastructure that keeps operational data inside the organization.

Key decisions

  • Citations over confidence. In an industrial setting, an unverifiable answer is a liability. Every claim is traceable to a source document.
  • Scope, then scale. The first release targets a bounded set of high-value document families rather than "everything" — it's easier to make a narrow copilot trustworthy than a broad one.
  • Evaluation gates. Retrieval and answer quality are scored against a labeled evaluation set before anyone in operations sees the tool.

Status

Concept case study — representative of the enterprise knowledge-copilot engagements XAI delivers, published as part of the studio's portfolio.