in-house-llm-data-analytics-assistant

The In-House LLM Data Analytics Assistant is a local, privacy-focused LLM analytics prototype that converts natural-language business questions into deterministic data analysis, retrieval-augmented generation (RAG), semantic category discovery, visualizations, and evidence-grounded explanations. It is built with Llama 3.1, Ollama, Python, Pandas, ChromaDB, Sentence Transformers, Altair, and Streamlit. The system uses Llama 3.1 as the natural-language intelligence layer while keeping authoritative analytical computation in deterministic tooling. Requests are routed across conversational, analytical, semantic-retrieval, and categorical-discovery workflows. For semantic questions, ChromaDB retrieves relevant records that are supplied to Llama as controlled context for grounded generation. The core engineering principle is to use the LLM for language understanding, intent interpretation, semantic reasoning, and explanation—while deterministic analytical tools perform authoritative calculations. The prototype was developed and tested against a synthetic structured dataset containing 30,000 client records, allowing the concept to be demonstrated without exposing confidential or production data. The architecture deliberately separates the data layer, deterministic computation layer, and LLM intelligence layer. The complete structured dataset is maintained in a Pandas DataFrame for deterministic analytics, while a separate vectorized representation of record-level content is stored in ChromaDB for semantic retrieval. The complete source dataset is not passed directly to Llama 3.1; instead, the model receives controlled context appropriate to the selected workflow.