RASS

Retrieval-Augmented Semantic Search (RASS)

Medical Informatics Engineering · Software Development Intern · May–August 2025

RASS lets engineers search more than 3,000 Redmine issues by meaning instead of keywords and get answers with citations back to the source issues. I built and containerized the first prototype and grew it into a multi-service platform.

What I built

• A pipeline that ingests documents, embeds them, retrieves with hybrid search in OpenSearch, and reranks the results. An LLM plans the query before retrieval.
• Separate services for asynchronous ingestion and for query serving, written in Node.js with Redis and PostgreSQL and deployed with Docker.
• REST, MCP, and SSE interfaces with structured citations, JWT and API-key authentication, and per-user security filters, so both people and AI agents can query it.
• Benchmarks for relevance, grounding, and latency across retriever and model settings. Context relevance stayed above 85% in RAGAS and TruLens checks that run before each deployment.
• Load tests at 25 concurrent users with responses under 500 ms, plus OpenTelemetry tracing so retrieval and service failures are easy to diagnose.

Stack

Node.js, Docker, OpenSearch, Redis, PostgreSQL, RAGAS, TruLens, OpenTelemetry

Links

• RASS / CoRAG repository