Ask VEPai. Trained chatbot interface for Ensembl VEP web

Ensembl VEP's web interface offers dozens of configuration options for variant annotation, overwhelming new users and generating recurring helpdesk queries. Ask VEPai is a locally-hosted, open-source AI assistant that translates natural-language descriptions into recommended VEP configurations with justifications. The system uses Retrieval-Augmented Generation (RAG) grounded in a curated knowledge base, with a deterministic constraint checker that catches species violations and option conflicts the model misses. Beyond the project brief's scope of option labelling, training data, and a prototype model, this proposal adds four components: (1) the constraint checker, empirically validated across three model sizes, (2) a quantitative evaluation framework with leave-one-out methodology and multi-run statistical replication, (3) a structured JSON schema mapping recommendations directly to VEP web form sections for "click to apply" integration, and (4) an interpretability layer where every recommendation includes a source citation traceable to the knowledge base. A working demo validated across Qwen 2.5 3B/7B/14B shows +19–30% Enable F1 from the knowledge base, with all code and results published. Core deliverables: expanded KB (~55 options), 15–20 gold-standard examples, RAG pipeline, constraint checker, JSON schema, evaluation report, and documentation. My availability (~400 hours) supports extending to a 350-hour Large project, adding QLoRA fine-tuning, attribution testing, FastAPI wrapper, and VEP output explainer.