LLM Integration

The Local AI section covered what the AI does. This section explains how it works under the hood. Quira's AI capabilities are powered by llama.cpp (cross-platform) and MLX (Apple Silicon). The model is downloaded on first run and stored locally.

How NL Query works internally

When you submit a natural language query, Quira uses a RAG (Retrieval-Augmented Generation) pipeline over your Context Graph:

1. Query embedding EYour question is converted to a 384-dimensional embedding vector.
2. Retrieval Esqlite-vec finds the most semantically similar nodes. FTS5 supplements with keyword matches.
3. Context assembly ERetrieved node summaries, entities, and metadata are assembled into a context window for the LLM.
4. Generation with citations EThe LLM generates a response grounded in the retrieved context, with mandatory source citations.

AI capability latencies