Generative AI Sandbox (POC)
Retrieval-Augmented Generation, Agentic workflows, and LLM orchestration validations.
Cospas-Sarsat Search & Rescue "Ask My Docs" RAG System
Architected and engineered a production-grade, highly deterministic Retrieval-Augmented Generation (RAG) platform to automate query processing across critical search and rescue (SAR) technical documentation manuals. Deployed an advanced hybrid search strategy that unifies dense semantic embeddings with sparse keyword tokenization via Reciprocal Rank Fusion (RRF) directly mapped to cloud database structures. The system mitigates downstream large language model hallucinations through strict bracketed source citation boundaries and utilizes a continuous integration quality-control framework to intercept code updates breaking performance or accuracy constraints.
Metrics at a Glance:
Key Architectural Achievements:
- Cloud Hybrid Search Optimization: Configured a dual-vector collection inside a free-tier Qdrant Cloud cluster, mapping text chunks simultaneously using dense conceptual vectors (BGE-Small) and deduplicated sparse lexical indices (BERT) merged instantly via server-side Reciprocal Rank Fusion (RRF).
- Deterministic Citation Constraints: Implemented an engineered system prompt framework instructing a high-speed Llama 3.1 inference engine via Groq to end every factual claim with matching bracketed token identifiers, instantly failing with an explicit fallback error code if source text bounds are insufficient.
- Resilient Streaming Ingestion Pipeline: Developed a robust file-system scanner utilizing PyPDF that reads multiple technical manuals simultaneously, automatically applying exponential back-off chunk streaming in sets of 15 data packets to eliminate write timeouts and network payload limits.
- Automated Regression Testing Gate: Built a localized validation testing runner file that acts as an automated cloud evaluator, evaluating pipeline changes against a golden question set to ensure new modifications do not cause quality drops or latency regressions.
- Reactive Operator Chat Dashboard: Wrapped the entire production-grade RAG backend into a fast web browser interface using Streamlit, featuring asynchronous query state management, clear latency tracker readouts, and a session-clearing controller dashboard.