ripunjay← All work

[ 04 ] Agentic RAG · Local project

Zenic

A local, source-backed assistant for nutrition and exercise questions, validated adult-profile calculations, and educational plans.

The browser UI and API run locally. The video demonstrates the project; the repository provides setup and validation steps. Zenic is educational software, not a clinical decision tool.

Zenic — ZENIC ASSISTANT — EVIDENCE-BASED HEALTH INTERFACE
ZENIC ASSISTANT — EVIDENCE-BASED HEALTH INTERFACE

Inside the build

Assistant workflows

Zenic answers nutrition and exercise questions against 10,201 bundled passages from NIH ODS, USDA, dietary guidelines, ISSN, and wger. It handles follow-ups, shows sources, and abstains when retrieval finds no usable evidence. Profile-validated requests can produce educational meal, workout, and weekly-summary PDFs; weekly demo data is synthetic.

Hybrid retrieval and evidence gate

BGE-small embeddings search local Chroma or production Qdrant. Reciprocal rank fusion combines vector search with BM25, then a cross-encoder reranks a bounded candidate set. Referential follow-ups are rewritten as standalone searches. An evidence gate excludes low-relevance passages before generation within a 16,000-character context budget.

Citations and their limit

Answers must cite supplied source IDs such as [1]. Missing or invalid IDs trigger abstention, and NIH citations include validated publisher links. The citation check confirms a cited source was supplied; it cannot prove every generated claim is supported by that source.

Validated calculations and plans

BMR, TDEE, and macro estimates are calculated in Python from a validated adult profile; the model explains the result. The LangGraph router sends calculation and plan requests through profile validation rather than asking the model to calculate them.

Browser UI and session behavior

A plain HTML, CSS, and JavaScript frontend talks to a Starlette API and streams progress while a turn runs. Browser sessions stay in memory for up to 30 minutes and do not survive a server restart. The repository does not include account authentication or a distributed rate limiter.

Validation and deployment

At the README’s last release check, 246 offline tests passed, along with ten live retrieval checks and a six-case RAG-versus-API routing sweep. Desktop and mobile flows and the Docker container were smoke-tested. RAGAS evaluation is optional; scores are not committed because they depend on the model, index, and provider state. The Docker image preloads models and runs as a non-root user.