MIRKO

Pharma QA Knowledge Copilot — RAG Over Regulated Documentation

A retrieval-augmented assistant for the quality-assurance department of a pharmaceutical manufacturer: it answers questions over SOPs, batch records and audit reports with citations, inside the company's private cloud.

Industry
Healthcare
Country
Ukraine
Team Size
5
Services
AI Integration, AI Solutions, IT Consulting
Technologies
Claude, Open-source LLMs, Python, pgvector, Elasticsearch, Azure, Kubernetes
Pharma QA Knowledge Copilot — RAG Over Regulated Documentation

Overview

About The Project

A pharmaceutical manufacturer maintains more than 18,000 controlled documents: standard operating procedures, batch records, deviation reports and audit findings. QA specialists spent hours locating the current version of a procedure and the history behind a deviation.

We delivered a knowledge copilot that answers questions over this corpus with citations to the exact document version, respects document access rights and runs entirely inside the client's private cloud.

The Challenge

Accuracy And AccessControl Were Non-Negotiable

In a GMP environment an answer based on an outdated SOP is worse than no answer. The system had to understand document versions, effective dates and supersession chains.

Documents carried different access levels, and the data could not leave the company network, which ruled out most hosted AI services.

Accuracy And Access Control Were Non-Negotiable

Our Solution

Context EngineeringFor A Regulated Corpus

We built an ingestion pipeline that parses the document management system, tracks versions and effective dates, and indexes only currently effective content by default, with history available on request. Retrieval combines semantic search with metadata filters for product, site and document type.

The assistant runs on a private Kubernetes cluster with an open-source model for retrieval and drafting and Claude through a private endpoint for complex reasoning when policy allows. Every answer shows its sources, and an evaluation set of 600 QA questions reviewed by the client's specialists gates each release.

Context Engineering For A Regulated Corpus

Key Features

What WeDelivered

The Impact

Results ThatMatter

QA specialists find the right procedure in seconds, with the evidence auditors ask for.

-85%

Time to locate a procedure

98%

Answers judged correct by QA

18,000+

Controlled documents indexed

0

Data leaving the private cloud