How Nova AI Learns Your Database Schema Automatically

See everything Query Streams can do beyond this guide Get Started Free

Supports

+more

Live Data

Replace exports with on-demand SQL

Excel Sheets

BI & AI Analytics

Ask Nova, get SQL + charts

Meet Nova

Database REST API

One key per partner. No credentials shared.

Build an API

+more

Automation

Scheduled sync to 6+ platforms

Explore

+more

API to SQL

Query APIs with SQL, no code

Explore

+more

AI Database MCP

Claude, Cursor, ChatGPT & Grok talk to your data

Connect AI

Query Streams is a secure, real-time database integration platform whose AI engine, Nova, learns your database schema automatically before generating a single SQL query. Learn more at QueryStreams.com and sign up for free to start asking your database questions in plain English.

Why Schema Understanding Matters for AI SQL

AI-powered SQL generation has a fundamental problem that most platforms ignore: the AI doesn’t know your database. Generic chatbots can write syntactically correct SQL, but they have zero knowledge of your actual table names, column types, data values, or relationships. Ask them to query your sales data and they’ll invent plausible-sounding column names like total_sales or customer_name that may not exist in your schema. The query looks right. It compiles. It returns either an error or, worse, wrong results from a different column that happened to match.

This is the hallucination problem applied to database schema analysis. The gap isn’t in the AI’s ability to write SQL syntax. Modern language models are competent SQL writers. The gap is in knowing what to write SQL against: which tables exist, which columns they contain, how tables relate to each other, what values appear in filter columns, and which SQL dialect the target database expects. Without this grounding, even the most sophisticated AI query builder is guessing. Schema Intelligence is how Nova eliminates that guessing entirely, creating an AI SQL assistant that operates on verified knowledge rather than probabilistic inference.

The 5-Phase Schema Intelligence Pipeline

Nova’s automatic schema learning runs a structured 5-phase pipeline the moment you connect a data source. Each phase builds on the previous one, progressively deepening the AI’s understanding of your database structure. The pipeline processes up to 10 tables concurrently for speed, supports 9 SQL dialects, and stores all learned knowledge persistently so the analysis only needs to run once per database unless your schema changes.

Phase 1: Schema Catalog Capture

The first phase connects to your database through the Query Streams Agent and reads the information schema catalog. This captures every table name, column name, data type, index definition, foreign key constraint, row count, and developer comment across all accessible schemas. The Agent runs on your own network and establishes a secure outbound-only connection, so no inbound ports, VPN tunnels, or firewall changes are required. Schema catalog capture is fast because it reads metadata, not data. A database with 200 tables typically completes this phase in under 3 seconds. The output is a complete structural map that Nova uses as the foundation for every subsequent phase and every SQL query it generates.

Phase 2: Data Sampling

Once Nova knows what tables and columns exist, it needs to understand what values they contain. Phase 2 sends a single ProfileTableForAI call per table through the Query Streams Agent. This specialized profiling operation replaces what would otherwise require 11 to 28 separate queries per table: sampling distinct values, detecting value distributions, identifying NULL rates, discovering enum patterns, and measuring cardinality, all in one network round-trip. Data sampling is what allows Nova to know that your state column uses two-letter abbreviations like “CA” instead of “California,” or that your status field contains “active”/”inactive” rather than 0/1. This prevents the most frustrating class of text-to-SQL errors: queries that are structurally correct but filter on values that don’t exist in your data.

Phase 3: Implicit Foreign Key Discovery

Many production databases have relationships between tables that aren’t declared as formal foreign key constraints. A customers.id column and an orders.customer_id column clearly relate to each other, but if no FK constraint exists in the schema, a basic catalog read won’t detect the connection. Phase 3 solves this by analyzing column names, data types, and value overlaps to identify undeclared relationships between tables. This implicit FK discovery is critical for complex natural language database queries that require multi-table JOINs. Without it, Nova would only be able to join tables with declared foreign keys, which in many real-world databases covers less than half of the actual relationships. By mapping both declared and implicit relationships, Nova can trace JOIN paths across your entire schema and generate accurate multi-table queries from questions like “show me orders with their customer names and product details.”

Phase 4: LLM Enrichment

Phase 4 passes the structural and profiling data to the AI for semantic enrichment. The LLM generates human-readable descriptions for every table and column, classifies semantic types (identifying that created_at is a creation timestamp, email is a contact field, sku is a product identifier), detects enum-like columns where a finite set of values repeats across rows, and maps tables to business domains such as sales, inventory, HR, or finance. This enrichment transforms raw technical metadata into contextual knowledge. When you ask “show me employee salaries by department,” Nova doesn’t just pattern-match the word “employee” to a table name. It understands which tables belong to the HR domain, which columns represent compensation data, and which column serves as the department grouper. Semantic classification is what separates a keyword-matching tool from an AI SQL assistant that genuinely understands your data model.

Phase 5: Result Processing

The final phase stores all learned knowledge and calculates coverage metrics. Nova tracks analysis completion at three levels: per-database (how many tables have been profiled), per-table (how many columns are enriched), and per-column (whether type classification and value profiling are complete). Result processing also includes contributor tracking. Every piece of learned knowledge is tagged as AI-generated or human-contributed. Team members can supplement the AI’s learning through a conversational training interface where Nova generates targeted questions about ambiguous columns or tables and humans answer in natural language. This collaborative approach combines the AI’s speed with human domain expertise, and the contributor tags ensure you always know the source of each annotation.

How Schema Intelligence Makes Nova Different

The gap between Nova and generic AI chatbots comes down to verified knowledge versus probabilistic guessing. When a general-purpose AI writes SQL for your database, it constructs queries based on statistical patterns in its training data. It might guess users when your table is actually app_users, or assume price when your column is unit_cost. These guesses are invisible to the user until the query fails or returns wrong results.

Nova eliminates guessing entirely. Every table, column, relationship, and data pattern referenced in a generated query comes from the Schema Intelligence knowledge base, verified against your actual database. Nova also enforces schema-qualified naming (using schema.table format like public.customers) and passes every generated query through SQLGlot syntax validation before execution. If a referenced column doesn’t match at a 60% similarity threshold (via rapidfuzz), Nova flags it and either corrects automatically or asks for clarification. This makes Nova not just an AI database schema reader but a SQL quality gate that catches errors before they reach your data.

The Schema Intelligence Interface

The Schema Intelligence interface, accessible at /nova/intelligence in the Query Streams portal, provides a visual dashboard of what Nova knows about each connected database. You can see per-table analysis status, coverage percentages for columns and relationships, and the source of each annotation (AI-generated or human-contributed). Per-entity analysis buttons let you trigger or re-run analysis on individual tables or columns on demand.

Four learning modes let you control how Schema Intelligence operates per database. Auto runs the full 5-phase pipeline automatically when a new data source is connected. Assisted notifies you before running the analysis and asks for confirmation. Manual lets you trigger analysis on individual tables using per-entity buttons. Off disables Schema Intelligence entirely for databases where you prefer to write SQL manually. The training interface is where human domain expertise meets AI speed: Nova generates targeted questions about gaps in its understanding, and team members answer conversationally. Each response is stored as human-contributed knowledge that supplements the automated analysis, which is particularly valuable for databases with non-obvious naming conventions or legacy schemas where column names don’t clearly indicate their purpose.

Don’t know SQL? Nova AI builds the query for you.

Try Nova

1

Describe what
data you need

Nova reads schema
& builds the SQL

Save & run from
Excel or Sheets

Help me create a query for last month’s revenue by product — I need it for my weekly sales report.

Found order_items and products in your schema. Here’s your query — run it to preview, then save it to your library and share it with your team.

SQL Results

Copy Run Save

SELECT
    p.product_name,
    SUM(oi.quantity * oi.unit_price) AS revenue
FROM   public.order_items oi
JOIN   public.products p ON oi.product_id = p.id
WHERE  oi.created_at >= DATE_TRUNC('month', NOW() - INTERVAL '1 month')
GROUP BY p.product_name
ORDER BY revenue DESC

Who Benefits from Automatic Schema Learning

Schema Intelligence isn’t limited to a single role. Any team that works with databases benefits from an AI that understands the data model before generating queries. Here’s how different roles gain from automatic database schema analysis.

Frequently Asked Questions

Get Started

Related guides: Text to SQL with Nova AI | All Features | Databases in Microsoft Excel | Databases in Google Sheets | API Connector

Category: Nova AI

Tags: Nova AI, Schema Intelligence, Database Schema, AI Analytics, Text to SQL, Machine Learning, Database Automation

Meta Description: Nova AI learns your database schema automatically. Better SQL from day one. Try free.

Author Bio:
The Query Streams Team comprises seasoned database experts, network security professionals, and enterprise networking veterans with over 25 years of combined experience in data management, analytics, and secure software development.