Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.delphina.ai/llms.txt

Use this file to discover all available pages before exploring further.

Ask a question about your data in plain English. The agent writes SQL, runs it against your warehouse, and returns the answer with data and charts.

​
A typical interaction

You type: 
What was our revenue last month by product category?
The agent responds with:
Revenue for February 2026 was $3.2M across all categories.
  • Enterprise SaaS — $1.8M (56%, +12% MoM)
  • Self-serve — $940K (29%, +3% MoM)
  • Professional Services — $460K (14%, -8% MoM)
Professional Services declined due to two large contracts completing in January. Enterprise SaaS growth was driven by 14 new accounts closed in the last 6 weeks.
Below the response, you can see the SQL the agent ran. If the agent made assumptions, the critic flags them as inline annotations. Follow up in the same chat — context carries forward: 
Show me the weekly trend for Enterprise SaaS over the last 3 months.

​
What happens behind the scenes

  1. Searches the knowledge base for relevant table schemas, metric definitions, and business rules.
  2. Queries your warehouse — writes and executes read-only SQL, then stages the results as local data files.
  3. Analyzes in a sandbox — runs Python and DuckDB against the staged data to transform, aggregate, and visualize. The sandbox is an isolated container with no internet access — your data never leaves the secure environment.
  4. Returns results — plain-English summary, data tables, charts, and the SQL used.
  5. Gets reviewed by a background critic agent that checks for missing knowledge or unjustified assumptions.
This two-step approach — SQL to stage, Python to analyze — means the agent can do things pure SQL can’t: statistical tests, complex reshaping, custom visualizations, and iterative exploration where each step builds on the last.

​
Types of questions

Metrics — “What was revenue last month?” “How many new customers in Q4?” Trends — “Daily active users for the past 30 days.” “Monthly growth rate this year.” Breakdowns — “Top 10 products by sales.” “Conversion rate by region.” Combined filters — “Revenue by category in the US for Q4, excluding test accounts.” Charts — “Bar chart of sales by region.” The agent picks a chart type, or you can specify one. Exploring your data — “What tables do you know about?” “What metrics are documented for sales?”

​
Using your knowledge base

The agent is significantly more accurate when it can draw on documented knowledge. If you’ve defined “MRR” in your knowledge base, the agent uses the documented SQL definition automatically — no guessing about which table or column to use. For metrics you use repeatedly, document them with /knowledge for consistency. See Maintenance.

​
SQL mode

Type /sql followed by a query to run it directly against your warehouse, bypassing the agent entirely: 
/sql SELECT date, revenue FROM analytics.daily_revenue
WHERE date >= '2024-01-01' ORDER BY date LIMIT 30
Your SQL runs exactly as written with read-only access. Useful for verifying the agent’s SQL, testing metric definitions before documenting them, or running queries you’ve already written.

​
File uploads

Drag and drop files into the chat to give the agent additional context:
  • Tabular data (CSV, TSV, Excel) — the agent can filter, aggregate, join with warehouse data, and create charts.
  • Documents (PDF, TXT, Markdown) — the agent reads the text and can answer questions about it.
  • Images (PNG, JPG, GIF, WebP) — the agent can describe charts, read text in screenshots, or reproduce visuals with live data.
To make files permanently available across all chats, upload them at Context > Sources > File Uploads.

​
Collaboration

All chats are visible to everyone in your workspace. Send a chat URL to share an analysis. Branching. Click Branch to create an independent copy of any conversation. Continue with your own follow-ups without affecting the original. Branch from a specific message to explore a different direction from midway through. Bookmarks. Click the bookmark icon on any response to save it for quick reference. View all bookmarks in the sidebar. Bookmarks are workspace-scoped — all members can see them.

​
Other modes

SituationUse
Multi-step investigation with a written report/research
Document a table, metric, or business rule/knowledge
Build your knowledge base from scratchAutomated Onboarding