What it does AVA is a TypeScript framework that ingests data from CSVs, JSON, URLs, or even raw text, then lets you query it in plain English. An LLM generates analysis code or SQL, summarizes results back into prose, and optionally spits out chart recommendations. It runs in both browser and Node.js, switching to SQLite or IndexedDB when your dataset crosses a configurable size threshold.

The interesting bit The architecture is deliberately modular—data loading, metadata extraction, analysis, LLM summarization, and visualization are separate stages—so you can yank out the chart generator or swap in a different model without rewiring the whole pipeline. The README also notes this is an experimental branch, which is refreshingly honest for a project claiming “AI-native” status.

Key highlights

• Accepts CSV strings, file objects, JSON, URLs, and unstructured text (with LLM parsing)
• Auto-suggests analysis queries ranked by relevance score and reasoning
• Runtime-aware storage: in-memory JS helpers under ~10KB, SQLite/IndexedDB above
• Optional visualization module detects intent and recommends chart syntax
• Ships with an llms.txt file for context-window priming

Caveats

• The 10KB default threshold for switching to SQLite/IndexedDB seems aggressively low for modern hardware; the README itself warns browser users to move large workloads server-side
• Only one LLM example (ling-1t) is shown; actual model support and token costs are unspecified
• Experimental branch status means APIs may shift

Verdict Worth a look if you’re building agentic dashboards or internal data tools and want to prototype fast without hand-rolling SQL generators. Skip it if you need battle-tested, deterministic analytics—this is glue code around LLMs, and it behaves like it.