Enterprise data agent
Ask your whole
data estate.
DelphiAI builds a semantic layer across your warehouse and your documents, then answers in plain English — with the SQL and the sources attached.
One question
Plain English in. Real SQL out.
No dashboard to find, no column names to learn. The agent picks the tables, writes the query and shows it to you — every answer arrives with the statement that produced it.
Semantic layer
It learns what your tables mean.
Point it at a connection and it reads the schema, infers the joins, names the metrics and writes the glossary. Your warehouse stops being a list of tables and becomes a map.
Structured and not
Numbers from tables. Reasons from documents.
Decks, PDFs, spreadsheets and images are parsed, chunked and linked to the tables they talk about — so “why is paid search ROI low” draws on the figure and the strategy that set it.
Architecture
Five layers, one answer.
From the systems your data already lives in, through a secured connection and a semantic layer, into memory, past an evaluator, and out as an answer or an action.
- 01 Enterprise data layer: Warehouses, documents, Slack channels and email — read where they live. Includes Structured databases, Documents, Slack channels, Email.
- 02 Context & semantic layer: What the tables mean: metrics, joins, a glossary, and documents linked to tables. Includes Metrics & glossary, Joins. Secured connection: Encrypted credentials, a read-only role, and every query grant-checked.
- 03 Context layer: Memory first; retrieval from the data layer when memory doesn't hold the answer. Includes Agent memory, Conversation history, Episodic memory, Semantic context.
- 04 Summarise & check: The draft is summarised, then scored by online evals before anyone sees it. Includes Summarise, Online evals.
- 05 Answer or action: A concrete answer with its SQL and sources — or the action it calls for.
Governed
Read-only, and only what they may read.
Every statement is parsed before it runs: one SELECT, no writes, a LIMIT injected, and each referenced table checked against the asker's grants. A layer never unlocks a table it was not granted.
Answers you can read
Shaped like a finding, not a transcript.
A figure, then the few lines that explain it, then the chart — rendered from a constrained spec the agent fills rather than code it writes. Nothing to clean up before you forward it.
DelphiAI is built to unlock enterprise potential for companies of every size to scale exponentially.
Context. One semantic layer over your warehouse and your documents.
Intelligence. A capable model for the answer, a fast one for the bulk work.
Permissions. Grants checked on every query, from the first one.
Efficiency. The layer narrows the question, so the prompt stays small.
What you get
Turn company knowledge into actions with DelphiAI.
Actionable insights
What happened, why, and what to do next — with the figures that prove it.
Easy integration with your company data and systems
Snowflake, Redshift, S3 and Google Drive, read where they live.
Your preferred LLM, to suit your needs
Anthropic or any OpenAI-compatible model, chosen in settings.
Enterprise-grade security.
People only see the tables they were granted.
Full observability.
Every question, query, row count and latency, audited.
Governed by design.
Read-only access, one checked SELECT at a time.
Built for the enterprise from day one. An answer only counts when it is secure, explainable and within what the asker may see.
See how access worksFrom question to action — one platform for your data.
Explore the platformConnect knowledge. Warehouses, buckets and drives, read where they live.
Understand context. Turn scattered tables and decks into one decision.
- Find the campaigns whose ROAS fell in Q3
- Read what the Q3 strategy deck says about them
- Summarise what changed, with the figures
Take action. Move from the answer to what happens next.
- 01search_semantic_layer'campaign roas q3'
- 02run_sqlSELECT campaign, roas FROM …
- 03search_documents'q3 strategy paid search'
- 04render_chart{ type: 'bar' }
Take it anywhere. Every answer carries its SQL, its chart and its sources, ready to paste where the work happens.
Get started
Connect a warehouse.
Ask a question.
Snowflake or Redshift, S3 or Google Drive. The first semantic layer builds itself while you watch.