Resources

Prove it on your data before you commit.

A ReefLink evaluation runs inside your environment, against your sources, with success criteria agreed up front. Here is how it works and what it answers.

Proof of concept

Federation first. Then AI.

Natural-language answers are only useful if the metadata, security, SQL correctness and performance underneath are controlled. The proof of concept validates them in that order — typically in three to four weeks once infrastructure and access are ready.

PhaseStageIndicativeActivitiesOutput
00Discovery & approval2–3 days

Confirm datasets, formats, network, security, AI deployment option, users and benchmark queries.

Approved scope

01Infrastructure readiness2–3 days

Provision hosts and containers; configure DNS, firewall, storage, certificates and service accounts.

Host readiness checklist

02Platform deployment2–3 days

Deploy coordinator and workers; configure memory, discovery, logging and restart behaviour.

Healthy cluster

03Source federation3–5 days

Configure catalogs; validate metadata, data types, connectivity and pushdown with query plans.

Federation evidence

04AI query layer4–6 days

Build metadata context, question-to-SQL workflow, validation and guardrails.

Controlled NL query flow

05Functional & performance tests4–5 days

Run agreed SQL, federation, concurrency, failure and AI accuracy test suites; capture metrics.

Benchmarks + accuracy scores

06Review & recommendation2 days

Document findings, gaps, risks, production changes and capacity assumptions.

Go / no-go report

Exit criteria

What the evaluation proves

  • The cluster stays healthy under functional and load tests; losing and re-adding a worker is demonstrated.
  • Approved datasets are discoverable and queryable through the platform.
  • Cross-source joins and aggregations complete, with documented query plans.
  • Pushdown is verified wherever the connector and source support it.
  • Baseline and optimized timings are captured for small, medium and large queries.
  • A curated set of 30–50 business questions is scored for SQL and answer correctness.
  • Generated SQL is read-only, restricted to approved objects, and rejected when validation fails.
  • Users cannot reach non-approved objects through SQL or the AI interface.

We don't promise latency numbers before your workloads are baselined. The evaluation establishes them.

What we need from you

Prerequisites

  • Linux hosts with container runtime, stable DNS and synchronized time.
  • Network routes from the cluster to the source systems.
  • Read-only service accounts scoped to representative datasets.
  • TLS certificates or enterprise CA trust where required.
  • A decision on AI deployment: approved managed model or self-hosted.
  • A small group of technical and business users to test and score results.

Documentation

Documentation

Product documentation — installation, connector configuration, policy administration and the AI layer — is provided to customers with each deployment.

Product

ReefLink documentation

Available to customers and evaluation teams. Ask your ReefLink contact for access.

Request access

Query engine

Trino documentation

SQL reference, connectors and administration for the open-source engine ReefLink is built on.

trino.io/docs

Governance

Apache Ranger

Policy model and administration for the access-control framework used by ReefLink.

ranger.apache.org

FAQ

Common questions

Does ReefLink copy our data?

No. Queries run against your sources in place through connectors. ReefLink stores metadata, query history and results you choose to keep — not copies of your tables.

Where does it run?

In your data centre or private cloud, deployed as containers. Your data does not leave your environment.

Which AI model does it use?

Your choice: an approved managed model API, or a self-hosted model on dedicated infrastructure for air-gapped environments. Only metadata and the question are sent to the model — never raw tables or credentials.

Can the AI change data in our systems?

No. Generated SQL is restricted to SELECT and EXPLAIN, validated before execution, and limited to approved objects. Source connections use read-only service accounts.

How do permissions work for AI questions?

Exactly as for typed SQL. The query runs as the user who asked, with the same Apache Ranger policies — column restrictions, masking and row filters apply.

Which sources are supported?

More than 30 source types including Teradata, Oracle, PostgreSQL, SQL Server, Snowflake, Iceberg, Delta Lake, Kafka and object storage. See the connector list. We also build custom connectors.

Is ReefLink a BI tool?

No. BI tools visualize prepared data. ReefLink is the layer beneath and around them: it connects and governs access to distributed data, investigates it, and produces findings. Your existing BI tools can connect to it over JDBC.

How long does an evaluation take?

Typically three to four weeks once hosts, network access and representative datasets are available.

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.