ReefLink
Sign In Request a Demo

AI

AI that works from your data, not around it.

ReefLink's AI is a governed layer on top of the query engine. It interprets a question using your metadata and business terms, writes SQL, validates it, and runs it under the asking user's permissions. Every answer shows where it came from.

Request a Demo See the guardrails

How a question runs

Question → context → SQL → validation → execution → result.

Each step is visible and logged. Run the example to see what happens between asking and answering.

Ask your data / signed in as ops_analystread-only

Example workflow. In a deployment, questions run against your approved catalogs.

  1. Question
  2. Context
  3. SQL
  4. Validate
  5. Execute
  6. Result

Press Run question to step through the pipeline.

Capabilities

Practical workflows, not a chatbot.

  1. 01

    Ask across systems

    Join inference from plain English, including questions that span several databases.

    • Uses known keys and relationships from metadata
    • One prompt, several catalogs
  2. 02

    Speak your business language

    Business glossary terms map to real columns and definitions.

    • Glossaries from Collibra, DataHub or Apache Atlas
    • Relative dates: "last quarter", "YTD", "this week"
  3. 03

    Summarize and relate

    Condense large collections of records and documents, and find how entities connect.

    • Every summary point links to its source
    • Relationship paths across systems
  4. 04

    Explain and diagnose

    Queries are explained in plain language; failed queries come back with a diagnosis and a corrected suggestion.

    • Readable explanation of generated SQL
    • Error diagnosis with proposed fix
  5. 05

    Draft intelligence briefs

    Turn a workspace of findings into a brief — marked as a draft for analyst review.

    • Situation, impact, recommendation
    • Confidence and gaps stated explicitly

Guardrails

A governed SQL layer — not an autonomous agent.

The model proposes. Deterministic checks decide what runs. Your policies decide what anyone can see.

ControlWhat it enforces
Read-onlyOnly SELECT and EXPLAIN are allowed. DDL, DML and administrative statements are blocked.
Allow-listGenerated SQL is restricted to approved catalogs, schemas and tables.
Permission-aware contextThe model is only told about objects the user is allowed to query.
SQL validationQueries are parsed before execution. Multiple statements and prohibited functions or patterns are rejected.
Query limitsRow limits, timeouts and resource-group constraints apply to every generated query.
No credentialsThe model never receives database passwords or secret values.
Prompt-injection resistanceDatabase content and user text are treated as untrusted and cannot override system policy.
Human visibilityGenerated SQL is shown before or alongside results so reviewers can check it.
AuditabilityPrompt, selected context, generated SQL, execution status and feedback are recorded, subject to your retention policy.

Model deployment

Your model, your boundary.

Only metadata and the question are sent to the model — never raw tables. Choose where inference runs.

Option A

Managed model API

When outbound access is permitted and data handling is approved. Fastest to set up; send metadata and question context only.

Option B

Self-hosted model

For air-gapped or data-residency requirements. Runs on dedicated inference infrastructure, separate from query workers so it does not compete for their CPU and memory.

Turn your data into intelligence.

See ReefLink run against your own sources — federated queries, governed access, and AI analysis that shows its evidence — inside your environment.

Request a Demo Talk to an Expert