AI tools dump your entire schema into a prompt and pray
Most NL-to-SQL products take your question, paste the full database schema into an LLM, and hope it picks the right tables. With 30 tables it works. With 300 it hallucinates. With multiple sources it falls apart entirely.
The engine understands before it queries
When you connect a source, Datural doesn't just read the schema — it builds a knowledge graph of what your data means. Relationships, business terms, naming patterns. When you ask a question, the engine retrieves only the tables that matter and hands the LLM a focused, precise context.
Instant understanding
Connect a database and the engine maps every table, column, and relationship in seconds. It generates human-readable descriptions so "acct_ltv_q4" becomes "account lifetime value, quarterly."
Semantic search, not string matching
Ask about "revenue" and the engine finds your "order_total" column. Ask about "churn" and it surfaces "subscription_cancelled_at." It understands meaning, not just names.
Cross-source intelligence
Users in Postgres. Events in MongoDB. Metrics in Databricks. The engine discovers that these sources share concepts and builds bridges between them automatically.
Focused context, not schema dumps
A 200-column database gets distilled to the 5 tables that matter for your question. The LLM sees exactly what it needs — nothing more, nothing less. Hallucinations drop dramatically.
Gets smarter with every question
Confirm a query is correct and the system remembers. Next time someone asks a similar question — across your whole team — the answer comes back instantly, without touching the LLM.
Transparent and auditable
Every answer traces back to specific tables, joins, and SQL. You see exactly which sources were used, how they connected, and why. No black boxes.
Under the hood
Connect and harvest
Point Datural at your databases. The engine crawls the schema, discovers relationships — explicit and implicit — and builds a knowledge graph with semantic descriptions for every entity.
Retrieve with precision
When you ask a question, the engine finds the exact tables needed through semantic matching and graph traversal. It constructs the minimal context — typically 3 to 7 tables out of hundreds — and hands it to the LLM.
Execute across sources
The LLM generates queries in each source's native language. Postgres gets SQL. MongoDB gets aggregation pipelines. Results are normalized and merged into a single, unified answer.
Learn and harden
Every confirmed answer strengthens the graph. Verified queries become cached routes. Business terms solidify into definitions. The engine gets faster, cheaper, and more accurate over time.
Schema dump vs. semantic engine
| Capability | Typical NL-to-SQL | Datural |
|---|---|---|
| Context strategy | Entire schema in prompt | Graph-selected subset |
| Accuracy at scale | Degrades past ~30 tables | Scales to hundreds |
| Cross-source queries | ✕ | Automatic |
| Learning from usage | ✕ | Every confirmed query |
| Repeat question cost | Full LLM call every time | Cached, instant |
| Relationship discovery | Explicit FKs only | Explicit + inferred + semantic |
| Transparency | Prompt in, SQL out | Full table + join trace |