nodejs-app

DB Agent (Node)

A Node.js/Express prototype of the same text-to-SQL agent as the root Streamlit app — built to compare how the UI feels on a real frontend stack (React + Tailwind + shadcn/ui) versus Streamlit's rerun-on-every-interaction model and widget styling. See the root README.md for background on the project.

Scope is intentionally narrower than the Python app in one respect: SQLite only, a single LLM (no model failover chain, no Databricks-native SQL backend). It does carry over the SQL repair-on-failure loop and the cross-platform contextual memory feature — see Notes below.

Two parts:

• server.js / memory.js — Express API (/api/schema, /api/config, /api/ask, /api/memories), also serves the built frontend
• web/ — React + Vite + Tailwind v4 + shadcn/ui frontend

Run it

Quick start with local Ollama (mirrors ../run_local.sh):

ollama pull qwen2.5-coder:7b   # or any model you prefer
cd nodejs-app
./run_local.sh

Checks Ollama is installed, the model is pulled, and the server is reachable before starting — same preflight checks as the Streamlit app's run_local.sh. Builds the frontend automatically on first run if web/dist doesn't exist yet. Pass a different model as $1: ./run_local.sh llama3.2.

Manual setup (any OpenAI-compatible endpoint):

cd nodejs-app
npm install
cp .env.example .env      # defaults to local Ollama; edit for OpenAI/Groq/etc.

npm run build              # builds web/dist
npm start

Open http://localhost:3001. Default DB_PATH is the self-contained ./data/demo.db bundled in this directory (so the app has no path dependency outside its own source tree — required for Databricks Apps, which deploys nodejs-app/ as an isolated source root). To instead share the root Python app's seeded DB during local dev, set DB_PATH=../data/demo.db (run the Python app once first, or python ../bootstrap.py, if that file doesn't exist yet).

Frontend dev mode (hot reload, instead of building): run npm run dev in web/ (http://localhost:5173) with server.js running separately — Vite proxies /api/* to http://localhost:3001 (see web/vite.config.ts).

Databricks Apps

This still fits the Databricks Apps Node.js runtime (React + Express are both documented-supported) — the deployment shape is the same generic "process bound to a port," just with one extra build step (npm run build in web/) before npm start.

Notes

• Uses Node's built-in node:sqlite module (stable in Node ≥ 22.5) — no native build step, unlike better-sqlite3.
• Mirrors ../core/sql_safety.py's safety rules and ../core/pipeline.py's SQL repair-on-failure loop (failed execution → error fed back to the LLM → re-validated before retrying) — same behavior, ported by hand since this is a separate runtime, not a shared library.
• Cross-platform contextual memory (memory.js, ported from ../core/memory.py): after each answered question, a second LLM call produces a redacted summary (never raw SQL/rows) written to a local JSONL store (data/memory_store.jsonl, gitignored). Other DBAGENT_ID instances pointed at the same store surface it as "Suggested from other agents" in the sidebar. Set DBAGENT_ID per instance (run_local.sh supports DBAGENT_ID=oltp-sqlserver ./run_local.sh, same as the Python app). Only the local JSONL backend is ported — the S3 Vectors backend (core/memory.py's S3VectorsBackend) is intentionally not, since it's still unverified against real AWS even on the Python side (repo issues #26–#30).
• No streaming, no auth — this is a UI comparison prototype, not a production alternative to the Streamlit app.