Strict schema validation, LLM fallback, robust graph builder, and GraphView refactor

[vishwateja231]

Motivation

• Harden LLM-to-SQL pipeline by removing permissive auto-corrections and enforcing exact schema names, while adding a fallback SQL generator to mitigate rate limits.
• Make graph construction data-driven so entities are discovered from arbitrary SAP column names and rows produce stable node/edge payloads.
• Simplify and modernize the frontend graph viewer for consistent styling, improved force layout, and more resilient handling of external graph payloads.

Description

• Replace permissive column/table aliasing with strict validation in schema_enforcer.py by disabling the column mapping and autocorrect_sql, tightening validate_sql_columns to check table-qualified refs and aliases, and keeping an authoritative SAP_TABLES list.
• Update SQL generation/validation flow in routers/query.py to remove mapping injection, stop auto-correct usage, add a single-retry UX for DB errors, change some error messages (e.g. return "Invalid column in query" when SQL generation fails), and add a Gemini-based SQL generation fallback via _gemini_sql_call when Groq is rate-limited.
• Rework graph_builder.py to be data-driven: introduce ENTITY_COLUMN_MAP, normalize values, extract entities per-row (_extract_entities_in_row), create nodes via _make_node, build edges from same-row entity order with a generic related label, and produce a fallback record node when no entities found.
• Refactor frontend GraphView.jsx to streamline UI and styling: update NODE_COLORS, remove the search box and error toast, add styleGraph to normalize incoming graph payloads, tune d3 force parameters, ensure theme-aware styles, and simplify ReactFlow panels and node/edge styling.

Testing

• Ran backend unit and integration checks including schema_enforcer and graph_builder smoke tests; checks passed locally with the existing test suite.
• Performed an end-to-end smoke run that exercised the LLM SQL generation fallback path and DB error retry logic, and observed successful fallback to Gemini when simulated rate limits occurred.
• Built the frontend (yarn build) and ran the dev server to verify GraphView renders styled nodes/edges and applies force layout; the build and manual render smoke checks succeeded.

---

Codex Task

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
dodge	Ready	Preview, Comment	Mar 25, 2026 5:40am