Semantic Search Engine for IT Infrastructure

The Legacy System

Our internal IT support platform relied on a rigid keyword-matching algorithm. This created a massive friction point: if a user searched for a 'login timeout' but the official documentation referenced 'session expiry,' the search returned zero results. Users were forced to manually dig through wiki trees or open support tickets for known issues.

Architectural Shift: The RAG Pipeline

I engineered a state-of-the-art Retrieval-Augmented Generation (RAG) pipeline to transition the platform from exact-match keyword mapping to contextual semantic understanding.

Query Processing & Vector Infrastructure

• Built a natural language processing layer using FastAPI to clean incoming user queries and normalize internal company jargon.
• Mapped text to vector embeddings utilizing OpenAI's text-embedding-3-large model.
• Migrated backend document storage to PostgreSQL, leveraging the pgVector extension. This allowed us to store high-dimensional embeddings and perform optimized Top-K similarity searches alongside standard relational metadata.

LLM Synthesis & Deployment

• Instead of just returning a list of links, the retrieved vector documents were passed to a secondary LLM layer.
• This layer synthesized the technical documents into a concise, human-readable summary, directly answering the user's question with cited internal sources.
• Deployed the API layer on AWS ECS for containerized, scalable performance, utilizing Celery for asynchronous background processing of new document embeddings.