Full-stack web application for GSI Helmholtzzentrum für Schwerionenforschung in Darmstadt. Transforms thousands of logbook entries—accumulated over years of complex equipment testing—into an intelligent, searchable knowledge base. Bachelor's thesis in Applied Computer Science; role: full-stack development, system architecture, ML pipeline design.
The Challenge
Traditional keyword-based search fails when terminology varies, synonyms differ ("cooldown procedure" vs "cryogenic preparation"), and answers require synthesising information scattered across multiple entries. Result: hours lost searching archives, knowledge loss during personnel transitions, repeated mistakes, delayed decision-making. The system addresses these by combining vector-based semantic search with large language models (RAG).
Solution and Architecture
- RAG pipeline: User query → embedding → vector search → context retrieval → LLM generation → cited answer. Answers are grounded in actual logbook content, traceable to source entries, and constrained to minimise hallucination.
- Three search modes: Semantic search (find by meaning); AI Q&A with citations (natural-language questions, synthesised answers with source references); Smart Search (vector similarity without LLM generation for faster browsing). Graceful degradation: falls back to keyword search when AI services are unavailable.
- Multi-dimensional filtering: By category, author, date range, test object, and more. User-defined context: Attach specific entries to constrain AI responses to selected sources.
Tech Stack
- Frontend: Next.js 14, React, TypeScript, Tailwind CSS, shadcn/ui, TanStack Query.
- Backend and data: PostgreSQL (Drizzle ORM), Pinecone (vector store), OpenAI (embeddings and LLM).
- Provider abstraction for local models (e.g. Ollama) or alternative providers; batch ingestion with idempotency for safe re-runs and incremental updates.
Results and Metrics
- Search quality: Precision@5 +60%, Recall@5 +31%, MRR +40% vs keyword search.
- Performance: Total RAG pipeline ~1.7s average, ~3.4s P95; query embedding, vector search, and DB fetch in the tens to low hundreds of ms; LLM generation dominates latency.
- Scale: Tested with 100,000+ entries; horizontal scaling via Pinecone serverless; incremental indexing.
Thesis Scope and Future
The thesis covers RAG foundations, system design and implementation, and evaluation of search quality and generated responses. Challenges addressed include domain-specific terminology (enriched document representation with metadata), long-document handling (truncation with preserved headers), and service reliability (graceful fallback). Future directions: local LLM integration (Ollama), real-time sync, custom embeddings, anomaly detection.